A method for predicting automotive user behavior based on a data model

Abstract

This application relates to the field of intelligent vehicle data processing, and discloses a method for predicting automotive user behavior based on a data model, including the following steps: Step 1, collect user multimodal data, which includes user driving trajectories, purchase records, in-vehicle videos, and voice commands; Step 2, perform standardization processing on 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 random matrix theory, eliminate the noise features in the covariance matrix and extract the features with higher signal strength; Step 5, adopt a feature interaction modeling method to optimize the cooperation relationship between signal features. The present invention adopts a technical solution combining multimodal data collection and standardization processing, and realizes the fusion of high-dimensional and diverse data under the same analysis framework by unifying the scales and distributions of different modal data.

Description

Technical Field

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

[0002] In the current field of intelligent vehicles, user behavior analysis and prediction have become an important direction for promoting personalized services and optimizing the user experience. With the popularization of intelligent in-vehicle devices, vehicles can collect multi-modal data, including but not limited to driving trajectories, user purchase records, in-vehicle videos, and voice interaction information. By mining and analyzing this multi-modal data, user behavior habits can be accurately predicted, providing support for the intelligent decision-making of vehicles. However, the high-dimensionality and heterogeneity of multi-modal data pose many challenges for analysis and modeling.

[0003] In the prior art, a user behavior prediction model is usually constructed using single-modal data. For example, only driving trajectory data is used for path planning, or purchase record data is used to analyze user preferences. Although these methods can achieve certain effects in specific scenarios, due to the failure to effectively combine the rich information of multi-modal data, the prediction accuracy of the model is limited. In addition, when dealing with high-dimensional multi-modal data, the prior art is easily interfered by inconsistent data scales and noise features, further reducing the stability and generalization ability of the model.

[0004] The prior art mainly has the following problems: In the process of processing multi-modal data, due to the significant differences in the scales and distributions of different modal data, directly using these data for modeling easily leads to uneven weight distribution among features, thereby affecting the model performance. In addition, when dealing with high-dimensional small-sample data, the unstable estimation problem of the covariance matrix will cause noise amplification in the feature extraction stage, thus reducing the reliability of the prediction model. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method for predicting automotive user behavior based on a data model, which solves the problems of inaccurate feature extraction and insufficient prediction accuracy of user behavior caused by inconsistent scales of multi-modal data and unstable estimation of the covariance matrix in the prior art.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for predicting automotive user behavior based on a data model, comprising the following steps:

[0007] Step 1, collect user multi-modal data, including user driving trajectories, purchase records, in-vehicle videos, and voice commands;

[0008] Step 2, perform standardization processing on the multi-modal data to construct a data matrix;

[0009] Step 3: Calculate the covariance matrix of the data matrix and improve the stability of the covariance matrix through an optimization algorithm;

[0010] Step 4: Based on the random matrix theory, eliminate the noise features in the covariance matrix and extract the features with higher signal strength;

[0011] Step 5: Optimize the cooperation relationship between signal features by using the feature interaction modeling method and adjust the feature weights through the attention mechanism;

[0012] Step 6: Input the optimized features into the user behavior prediction model to complete the prediction of user behavior;

[0013] Step 7: When new data arrives, dynamically update the covariance matrix and the user behavior prediction model, and adjust the prediction results in real time.

[0014] Preferably, the normalization processing of the multi-modal data includes normalizing the data of different modalities according to the mean and standard deviation of their feature dimensions, so that the data is comparable within the same dimension.

[0015] Preferably, the optimization of the covariance matrix is completed by 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.

[0016] Preferably, the steps of eliminating noise features by the random matrix theory include the following:

[0017] S1. Analyze the eigenvalue spectrum distribution of the covariance matrix;

[0018] S2. Determine the distribution range of noise eigenvalues according to the Marchenko-Pastur theorem;

[0019] S3. Eliminate the eigenvalues and their corresponding eigenvectors that fall within the noise range.

[0020] Preferably, the interactive modeling of the signal features is completed by constructing a quadratic function, and the quadratic function is composed of an interactive relationship matrix of feature vectors, linear weights, and bias terms.

[0021] Preferably, the user behavior prediction model adopts the Transformer framework, which includes an input layer, a self-attention layer, a feed-forward neural network layer, and an output layer, where the self-attention layer is used to capture the correlation between signal features.

[0022] Preferably, the dynamic update includes the following:

[0023] Step 1: Normalize the new data;

[0024] Step 2: Calculate the updated covariance matrix and perform weighted combination with the historical covariance matrix according to the time decay factor;

[0025] Step 3: Adjust the parameters of the prediction model based on the updated covariance matrix.

[0026] Preferably, the extraction of the signal features includes vectorized representation of audio and video data, and integrating it with structured data to form a unified data matrix.

[0027] Preferably, the optimization objective of the prediction model is to complete the adjustment of 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.

[0028] The present invention provides a method for predicting automotive user behavior based on a data model. It has the following beneficial effects:

[0029] 1. The present invention adopts a technical solution combining multi-modal data acquisition and standardized processing. By unifying the scales and distributions of different modal data, the fusion of high-dimensional and diverse data under the same analysis framework is achieved. Compared with the prior art that only models single-modal data, the present invention solves the problem of model performance degradation caused by inconsistent scales between multi-modal data, and significantly improves the accuracy and consistency of data analysis.

[0030] 2. The present invention introduces Ledoit-Wolf shrinkage estimation to optimize the covariance matrix, and improves the stability of the covariance matrix by using the weighted combination of the identity matrix. Different from the traditional covariance matrix calculation method that is prone to unstable eigenvalue distribution in the small sample high-dimensional scenario, the present invention effectively solves the deficiencies of overfitting and large estimation bias of the covariance matrix in high-dimensional data, and ensures the accuracy of subsequent feature extraction.

[0031] 3. The present invention analyzes the distribution range of the eigenvalues of the covariance matrix through random matrix theory, and the present invention 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 prone to noise interference, the solution of the present invention makes the extracted features more accurate, providing high-quality input for subsequent user behavior modeling.

[0032] 4. The present invention designs a dynamic update mechanism combining the time decay factor and incremental learning, which can update the covariance matrix and the parameters of the prediction model in real time. Compared with the prior art that requires retraining the model to adapt to new data, the present invention effectively solves the deficiencies of the model in real-time performance and computational efficiency, and significantly improves the adaptability and efficiency of user behavior prediction. Description of the Drawings

[0033] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed implementation manners

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] Please refer to the appendix Figure 1 The embodiments of the present invention provide a method for predicting automotive user behavior based on a data model, including the following steps:

[0036] Step 1: Collect multi-modal data of users, including user driving trajectories, purchase records, in-vehicle videos, and voice commands; Step 2: Perform standardization processing on the multi-modal data to construct a data matrix;

[0037] Step 3: Calculate the covariance matrix of the data matrix and improve the stability of the covariance matrix through an optimization algorithm;

[0038] Step 4: Based on random matrix theory, eliminate the noise features in the covariance matrix and extract the features with higher signal strength;

[0039] Step 5: Optimize the cooperation relationship between signal features by using a feature interaction modeling method and adjust the feature weights through an attention mechanism;

[0040] Step 6: Input the optimized features into the user behavior prediction model to complete the prediction of user behavior;

[0041] Step 7: When new data arrives, dynamically update the covariance matrix and the user behavior prediction model, and adjust the prediction results in real time.

[0042] The standardization processing of multi-modal data includes normalizing the data of different modalities according to the mean and standard deviation of their feature dimensions, so that the data is comparable within the same dimension.

[0043] The optimization of the covariance matrix is completed through 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.

[0044] The steps of eliminating noise features by random matrix theory include the following:

[0045] S1: Analyze the eigenvalue spectrum distribution of the covariance matrix;

[0046] S2. Determine the distribution range of the noise eigenvalues according to the Marchenko-Pastur theorem;

[0047] S3. Eliminate the eigenvalues and their corresponding eigenvectors that fall within the noise range.

[0048] The interactive modeling of signal features is completed by constructing a quadratic function, which consists of an interactive relationship matrix of eigenvectors, linear weights, and bias terms.

[0049] The attention mechanism adopts a multi-head attention mechanism to complete the importance ranking of signal features by calculating the correlation weights between signal feature vectors.

[0050] The user behavior prediction model adopts a Transformer framework, including an input layer, a self-attention layer, a feed-forward neural network layer, and an output layer, where the self-attention layer is used to capture the correlation between signal features.

[0051] The dynamic update includes the following:

[0052] Step 1. Standardize the new data;

[0053] Step 2. Calculate the updated covariance matrix and perform weighted combination with the historical covariance matrix according to the time decay factor;

[0054] Step 3. Adjust the parameters of the prediction model based on the updated covariance matrix.

[0055] The extraction of signal features includes the vectorized representation of audio and video data, and after integrating it with structured data, a unified data matrix is formed.

[0056] The optimization objective 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.

[0057] In the technical solution of the present invention, the acquisition of multi-modal data is the starting point of the entire user behavior prediction process, and its quality directly affects subsequent feature extraction, modeling, and prediction accuracy. Generally, multi-modal data includes, but is not limited to, information such as the driving trajectory of users, purchase records, in-vehicle videos, 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 spatio-temporal correlation of different modality data also needs to be comprehensively considered to ensure data consistency and integrity.

[0058] As an option, these multi-modal data can be collected in real time from multiple data sources and integrated into a unified storage structure. Specifically, the present invention designs a reasonable data acquisition mechanism to provide high-quality original input for subsequent data processing.

[0059] In a possible implementation, the driving trajectory data can be obtained through the vehicle's Global Positioning System (GPS) module. Specifically, the GPS module can record information such as the vehicle's timestamp, latitude and longitude 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 time point. To ensure the accuracy and integrity of the trajectory data, generally, the missing points or duplicate points that may occur during the acquisition process are preprocessed. For example, the missing points are filled in through an interpolation algorithm, and the duplicate points are removed through a deduplication algorithm.

[0060] As an improved method, the acquisition of trajectory data can also be combined with the data of the Inertial Measurement Unit (IMU). For example, the vehicle's accelerometer and gyroscope are used to supplement the driving trajectory information to form a more accurate trajectory data matrix.

[0061] The formula is expressed as follows: x i，j =[lat j ,lon j ,spd j ,acc j ,yaw j Where: x i，j is the trajectory feature at the j-th time point; lat j and lon j are the latitude and longitude at the j-th time point respectively; spd j is the speed of the vehicle at the j-th time point; acc j is the acceleration of the vehicle; yaw j is the yaw angle of the vehicle.

[0062] Specifically, the user purchase record data can be obtained by connecting to the database of the car service platform. Such data includes historical order information, transaction amount, transaction time, and product category generated by the user on the platform. These data have a structured characteristic and can be directly organized in matrix form.

[0063] 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 transformed into numerical features through natural language processing (NLP) methods, such as keyword extraction, sentiment analysis score, etc.

[0064] To unify the structured representation of these purchase records, they can be organized into the following data matrix: p i，k =[amt k ,cat k ,time k ,nlp k Where: p i，k is the feature of the k-th purchase record; amt kis the purchase amount; cat k is the numerical code for the product category; time k is the transaction timestamp; nlp k is the language feature score extracted from this transaction record.

[0065] As an option, the acquisition of in-vehicle video data can be achieved through a camera installed in the vehicle. Generally, the video data is segmented into a frame sequence, and features are extracted from each frame image through preprocessing, such as object detection, driving scene analysis, etc. Further, the video frame features are transformed into a fixed-length vector representation through dimensionality reduction technology to form a video feature matrix.

[0066] Voice data can be collected through an in-vehicle speech recognition system, such as voice commands when the user interacts with the vehicle. The collected voice data first extracts spectral features through frame segmentation and Fourier transform, and then applies a convolutional neural network (CNN) to generate high-dimensional voice features.

[0067] In a possible implementation, the video and voice features are integrated into a hybrid modality feature matrix, and the formula is as follows: v i，l =[vid l , spec l where: v i，l is the hybrid modality feature at the l-th time point; vid l is the feature vector of the video frame; spec l is the spectral feature vector of the voice signal.

[0068] In the design of the present invention, different modality data needs to be integrated into a unified data storage format. Specifically, the multi-modal data is aligned through timestamps to form a global feature matrix. The aligned data matrix can be expressed as follows: where: X is the integrated data matrix; N is the length of the time series; M is the total number of features, including all features of trajectories, purchase records, videos, and voices.

[0069] In a possible implementation, a distributed storage system can be used to store the integrated data for subsequent large-scale processing.

[0070] In the technical solution of the present invention, normalization processing is a key link for multi-modal data to enter the analysis stage. Its purpose is to make the data comparable within a unified dimension range by eliminating the scale differences of different modality data. Generally, the sources of multi-modal data are extensive, 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, and direct calculation may cause the weights of some modality features to be too large or too small, thus affecting the calculation of the subsequent covariance matrix and the modeling accuracy.

[0071] As an option, all modal data can be standardized through a normalization method and mapped to the same numerical range. Specifically, the standardization process usually combines the mean and standard deviation of each modal data to ensure that the eigenvalue is on the same distribution scale.

[0072] In this embodiment, the multi-modal data collected in step one is organized into a unified data matrix X, where where N represents the number of samples and M represents the total dimension of features. Each row of the matrix X represents a sample, and each column corresponds to a feature. Generally, each element x of the data matrix i，j is the original collected data, and there may be large differences in the numerical ranges of different modal features. For example, the longitude and latitude coordinates in the driving trajectory may have a value range between [0, 180], while the value range of the vehicle speed may be [0, 200]. This difference may cause specific modal data to dominate in subsequent analysis and interfere with the accuracy of the covariance matrix.

[0073] To eliminate the above problems, the present invention standardizes each column of the data matrix, and the specific formula is: where: x' i，j is the eigenvalue of the i-th row and j-th column after standardization; x i,j is the original eigenvalue before standardization; is the mean of the j-th column feature; is the standard deviation of the j-th column feature.

[0074] Generally, the mean μ j and the standard deviation σ j are calculated within the entire sample data range, so as to ensure the consistency of the standardized features of all samples.

[0075] In some embodiments, for discrete features (such as the commodity category code in the purchase record), the standardization process may not be applicable. As an option, one-hot encoding can be used to convert the discrete feature into a multi-dimensional sparse vector form. For example, cat j The encoding result can be expressed as: cat j = [0, 1, 0, 0, 1], where each dimension represents whether a certain commodity category is relevant to the current sample.

[0076] Specifically, the standardization process of unstructured data such as speech and video needs to combine specific modal features. For the spectral features of speech signals.

[0077] For video data, standardization can be performed by normalizing pixel values and reducing the dimension after feature extraction. For example, the pixel value pixi，j Process according to the following formula: where the pixel value range [0, 255] is compressed to [0, 1].

[0078] In a possible implementation, the normalization process of multimodal data follows the following steps: First, normalize all numerical features according to the mean and standard deviation; Second, perform one-hot encoding on discrete features for feature expansion; Finally, normalize unstructured modal features (such as videos, voices) within a specific modal range and integrate them into the data matrix X.

[0079] After the above processing, all modal data are mapped to the same numerical range, and finally a standardized data matrix X′ is formed for subsequent calculation of the covariance matrix.

[0080] As an improved method, the present invention can also perform noise removal on the standardized data matrix. For example, for possible outliers in trajectory data (such as sudden changes in speed or yaw angle), a threshold filtering method can be used for correction. For example, if the acceleration of a vehicle exceeds a reasonable range (such as 10m / s 2 ), then it is replaced with the mean acceleration of adjacent time points: In the technical solution of the present invention, the calculation and optimization of the covariance matrix are indispensable key steps in the user behavior prediction process. Generally, the standardized data matrix X contains the feature relationship information of multimodal data, while the covariance matrix can comprehensively 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 the scenario of high-dimensional small samples (i.e., M > N), 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.

[0081] Specifically, the present invention designs a covariance matrix optimization method based on Ledoit-Wolf shrinkage estimation. By combining the linear weight of the identity matrix, the stability of the covariance matrix is significantly improved, laying a solid foundation for subsequent feature extraction and modeling.

[0082] In this embodiment, based on the standardized data matrix obtained in step two, the covariance matrix is calculated. Each element of the covariance matrix represents the linear correlation between features, and its definition is as follows: where: represents the covariance matrix; is the standardized data matrix; N represents the number of samples; M represents the number of features.

[0083] In the above formula, XT Represents the transpose operation of the matrix X. Each element C of the matrix C i，j Defined as: where x k，i and x k，j represent the i-th and j-th eigenvalues of sample k respectively.

[0084] In general, the covariance matrix C is a symmetric matrix with diagonal elements C i，i Represents the variance of a single feature, the off-diagonal elements C i，j Represents the covariance between two features.

[0085] In actual scenarios, especially when N<M (that is, 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 appear 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.

[0086] As an option, the stability of the covariance matrix can be improved by introducing linear weighting of the identity matrix.

[0087] 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: LW =(1-ρ)C+ρIwhere: represents the optimized covariance matrix; represents the original covariance matrix; is the unit matrix; ρ∈[0, 1] represents the contraction strength.

[0088] In general, the value of ρ is determined by the number of samples and the number of features, and its calculation formula is: Where: Var(C i，i ) represents the variance of the diagonal elements of the covariance matrix; Represents the squares of the diagonal elements.

[0089] Specifically, the introduction of the shrinkage strength ρ can reduce the random fluctuation effect between features, thereby improving the numerical stability of the covariance matrix.

[0090] 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.

[0091] In a possible implementation, the PCA method can be used to reduce the dimensionality of the data matrix first, and then the covariance matrix C' after dimensionality reduction is optimized by shrinkage.

[0092] After the above optimization, the covariance matrix C LW has the following characteristics: the diagonal elements still represent the optimized eigenvalue variances; the off-diagonal elements represent the optimized eigen-covariances, reducing noise interference; the eigenvalue distribution of the covariance matrix is smoother, providing a reliable basis for subsequent noise removal and modeling.

[0093] Specifically, the present invention uses the Marchenko-Pastur theorem to analyze the eigenvalue distribution of the covariance matrix, and by identifying the range of noise eigenvalues, the corresponding eigenvectors are removed, so as to retain the part with significant signal characteristics. The optimized eigenmatrix provides reliable support for subsequent modeling and prediction.

[0094] In this embodiment, random matrix theory is used to analyze the eigenvalue spectrum distribution of the covariance matrix. Generally, the eigenvalue distribution of the covariance matrix can be described by the Marchenko-Pastur theorem. According to this theorem, when both the number of samples N and the feature dimension M tend to infinity and the ratio is constant, the eigenvalue distribution of the covariance matrix has the following range: Where: represents the ratio of the number of samples to the feature dimension; λ + and λ - are the upper and lower bounds of the eigenvalues respectively.

[0095] Specifically, when the eigenvalues of the covariance matrix are within the range of [λ - , λ + , these eigenvalues are generally considered to be caused by noise; while the eigenvalues outside this range are considered to have significant signal strength.

[0096] To achieve the removal of noise features, the following steps are designed in this embodiment:

[0097] First, by performing eigenvalue decomposition on the covariance matrix, the eigenvalues and eigenvectors of the covariance matrix are obtained. Generally, the covariance matrix C can be decomposed as: C = VΛV T Where: is the eigenvector matrix, and its column vector V i represents the eigenvector of the covariance matrix; Λ = diag(λ1, λ2,..., λ M ) is the eigenvalue diagonal matrix, where λ i represents the eigenvalue of the covariance matrix.

[0098] Subsequently, the upper and lower bounds λ of the eigenvalues are calculated according to the Marchenko-Pastur theorem ± . Specifically, for the number of samples N and the number of features M, the calculation formulas for the upper and lower bounds are as follows: After calculating λ ± , the eigenvalues within the range [λ - , λ + and their corresponding eigenvectors are removed. The optimized covariance matrix can be constructed by the following formula: Where: λ i represents the eigenvalues beyond the noise range; v i is the corresponding eigenvector.

[0099] In some embodiments, the process of noise removal can also be combined with the calculation of the feature contribution rate to ensure that the retained features after noise removal are comprehensive enough. For example, the feature contribution rate can be defined as the ratio of a certain eigenvalue to the sum of all eigenvalues: When the cumulative contribution rate ∑r i exceeds a certain threshold (such as 95%), the subsequent feature removal is stopped. This improved method can ensure the retention of signal features and further optimize the effect of noise removal.

[0100] In a possible implementation, λ ± can also be adjusted by introducing a regularization parameter to adapt to different data distributions. For example, when the noise in the data is high, a lower threshold of λ _ can be selected to remove more noise features.

[0101] After the above noise removal process, the dimension of the covariance matrix will be significantly reduced, and only the part with higher signal strength is retained. At this time, the signal feature matrix can be represented by the following formula: Where: represents the eigenvector corresponding to the signal eigenvalue ; k is the number of signal features.

[0102] In the technical solution of the present invention, the signal feature matrix W extracted through the foregoing steps mainly includes the part with significant signal features in the covariance matrix. However, there may be complex correlations and interaction relationships between these signal features. Modeling based on a single feature may ignore the synergistic effects between features and reduce the prediction accuracy. Generally, in order to more comprehensively utilize the signal features, the present invention designs a feature interaction modeling method to optimize the collaborative relationship between signal features.

[0103] Specifically, the present invention constructs a quadratic function to describe the interaction between signal features, and further combines the multi-head attention mechanism to optimize the weights of the importance of signal features. This method can accurately capture the relationships between features while highlighting the features highly relevant to the prediction task, thereby improving the modeling effect.

[0104] In this embodiment, each column W of the signal feature matrix W i represents the i-th signal feature vector extracted, and the dimension of the signal feature matrix W is where: N represents the number of samples; k represents the number of signal features.

[0105] To model the interaction relationship between signal features, the following quadratic function is constructed in this embodiment: f(W) = W T QW + b T W + c where:

[0106] is the interaction relationship matrix, representing the relationship strength between different signal features;

[0107] is the linear weight vector, representing the direct contribution of a single feature to the target;

[0108] is the bias term.

[0109] Generally, the interaction relationship matrix Q is a symmetric matrix, and each of its elements Q i，j represents the interaction strength between the features w i and w j . By minimizing the error function, the values of Q, b, and c can be optimized to accurately reflect the actual relationships between features.

[0110] In this embodiment, the 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 globally, thereby assigning more reasonable weights to the features. Specifically, the calculation formula of the multi-head attention mechanism is as follows: where: Q = W·W T is the query matrix, representing the correlation between features; K = W is the key matrix; V = W is the value matrix; d k is the feature dimension.

[0111] As an improvement method, in this embodiment, a weighted average is also performed on the output of the attention mechanism to further highlight the features related to target prediction. The weighted feature matrix can be expressed as: W′ = αW where α is the attention weight vector, and each of its elements is calculated by the softmax function.

[0112] In some embodiments, the quadratic function for feature interaction modeling can be extended to a form with a regularization term to improve the generalization ability of the model. For example, a

[0113] regularization term can be added to make the optimization objective become: where: represents the true value of the prediction target; is the Frobenius norm of the interaction relationship matrix Q; λ is the regularization coefficient used to control the strength of regularization.

[0114] In addition, during the weight optimization process, the calculation method of α can be adjusted by combining the characteristics of different modal features. For example, for video modal features, the importance of features at different times can be highlighted by introducing a time series attention mechanism.

[0115] After the above-mentioned 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 at the same time highlight the features highly relevant to the prediction task. Specifically, the optimized feature matrix can be expressed as: W′ = Attention(Q, K, V)·W + b T W + c In the technical solution of the present invention, the optimized signal feature matrix W′ is obtained through the foregoing steps. This matrix retains the part with significant signal features in the data and further improves the expression ability of the features through feature interaction modeling and attention weight optimization. On this basis, to achieve accurate prediction of user behavior, the present invention designs a behavior prediction model based on deep learning. Generally, the optimized feature matrix W′ is used as the input of the model, and the data features are further extracted and modeled through a multi-layer structure, and finally the prediction result of user behavior is generated.

[0116] As an option, the present invention adopts a Transformer framework to construct a prediction model, combines the expression ability of the global attention mechanism and the feed-forward neural network, and realizes accurate modeling and behavior prediction of high-dimensional multi-modal data.

[0117] In this embodiment, the user behavior prediction model consists of a multi-layer network structure, specifically including an input layer, a self-attention layer, a feed-forward neural network layer, and an output layer. The core of the model is to capture the high-order relationship between optimized features through the self-attention mechanism, thereby improving the prediction accuracy.

[0118] Design of the input layer Generally, the input layer is used to receive the optimized signal feature matrix W′. The shape of the input matrix is where: N is the number of samples; k is the number of signal features.

[0119] In the input layer, positional encoding (PositionalEncoding) is added to each feature to preserve the context information of the feature. The positional encoding can be expressed as: where: i represents the position index of the feature; j represents the dimension index of the feature.

[0120] After adding the positional encoding to the input matrix W′, the final embedding matrix W is generated embed , which is used as the input for the subsequent network layers.

[0121] Generally, to enhance the model's ability to capture different feature interaction patterns, the self-attention mechanism is extended to the multi-head attention mechanism. Its output can be expressed as: MultiHead(Q, K, V) = Concat(head1, head2,..., head h ) · W O where: head i = Attention(Q i , K i , V i ) represents the calculation result of the i-th attention head; h is the number of attention heads; W O is the linear transformation weight matrix.

[0122] Based on the output of the self-attention layer, the feed-forward neural network layer further extracts non-linear features. Specifically, the feed-forward neural network layer contains two fully connected layers, and its expression form is as follows: FFN(x) = ReLU(x · W1 + b1) · W2 + b2 where: x is the output of the previous layer; W1, W2 are weight matrices; b1, b2 are bias vectors; ReLU(·) is the activation function, representing the rectified linear unit.

[0123] Generally, the output layer is responsible for generating the prediction result of the 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 the user behavior. The specific formula is as follows: where: represents the prediction value of the i-th sample; z i represents the linear transformation output of the last layer of the model.

[0124] In some embodiments, the training objective of the model can be adjusted according to specific task requirements. For example:

[0125] For binary classification tasks, the cross-entropy loss function can be used as the optimization objective: where y i is the true label, is the model prediction value.

[0126] For multi-classification tasks, it can be extended to the Softmax loss function.

[0127] 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.

[0128] 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.

[0129] 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.

[0130] 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: new =βC old +(1-β)C urrent Where: β∈[0,1] is the time decay factor, which controls the weight of new and old data; C old is the historical covariance matrix; Represents the covariance matrix of the new data; M new is the number of samples of new data.

[0131] In general, the value of the time decay factor β can be adjusted according to the specific application scenario. For example, in a scenario with high real-time requirements, a smaller β can be set to increase the weight of new data.

[0132] As an option, β can be dynamically adjusted to automatically optimize the weight distribution based on the difference between new data and historical data. Specifically, the dynamic adjustment formula can be expressed as: where ||·|| F Represents the Frobenius norm of the matrix, which is used to measure the difference between matrices.

[0133] After the covariance matrix is updated, the extraction method of the signal feature matrix W also needs to be adjusted accordingly. In general, based on the new covariance matrix C new Recalculate the eigenvalue distribution and remove the noise features. The specific steps include:

[0134] First, according to the Marchenko-Pastur theorem, the updated eigenvalue upper and lower bounds λ are calculated. ± : in is the number of new samples, and M is the feature dimension.

[0135] Then, remove the - ,λ + ] and only retain the signal features.

[0136] Finally, reconstruct the signal feature matrix W new , as the dynamically updated feature input.

[0137] 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.

[0138] As a possible implementation method, the model parameters can be fine-tuned through the gradient update method. The specific formula is as follows: in:

[0139] θ old and θ new Respectively represent the parameters before and after the model update; η is the learning rate; is the gradient of the loss function on the new data.

[0140] 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 automotive 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; 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; 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; 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; 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; 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, wherein the self-attention layer is used to capture the correlation between signal features; 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.

2. The method for predicting automotive user behavior based on a data model according to claim 1, wherein, 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. A method for predicting automotive 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. Determine the distribution range of noise characteristic values according to the Marchenko-Pastur theorem; S3. Eliminate eigenvalues and their corresponding eigenvectors that fall within the noise range.

4. A method for predicting automotive 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.

5. A method for predicting automotive 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.

6. A method for predicting automotive 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