A method and system for information hierarchical classification and intelligent distribution for a travel information platform

Abstract

The application discloses a kind of information hierarchical classification and intelligent distribution method and system for travel information platform, it is related to machine learning technical field.The method comprises the following steps: acquiring multi-source heterogeneous time series data, inputting to multimodal feature embedding network, the multimodal feature embedding network carries out history contribution degree dynamic adjustment to time series feature, and after decoupling time series feature and spatial feature with mutual information minimization constraint, output fusion feature, the fusion feature is input to multi-label classification neural network, the parameters of the multi-label classification neural network are maintained by gradient orthogonal projection and branch momentum, and the classification result is obtained, and the classification result is input to strategy network as state vector.The application solves the gradient suppression problem under the label long tail distribution by orthogonal projection of tail category gradient to the direction decoupled with high-frequency category, and independently maintains the momentum cache with frequency self-adaption for each category, to ensure that low-frequency high-priority information is correctly divided into high level.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a method and system for information classification and intelligent distribution for travel information platforms. Background Technology

[0002] Travel information platforms are a core component of intelligent transportation systems. They integrate multi-source heterogeneous data, including location trajectories, motion states, and user interaction behaviors, to classify travel information across multiple dimensions, such as security level, content category, time attributes, and spatial range. Personalized information pushes are then implemented based on real-time user scenarios. With the rapid development of vehicle-to-everything (V2X) and mobile communication technologies, travel information platforms need to process massive amounts of heterogeneous data and make accurate distribution decisions in complex and dynamic environments. The core challenges are: how to extract stable representations from multi-source heterogeneous time-series data for multi-dimensional classification, and how to achieve dynamic adaptive distribution while satisfying multiple security constraints.

[0003] To address the issue of multi-dimensional classification and distribution of travel information, the invention patent with publication number CN116720140A discloses a method and system for generating vehicle network information distribution strategies based on deep belief networks. The method uses deep belief networks to extract features from vehicle network accident information and introduces a two-layer multi-label classification network at the top layer of the deep belief network. The distribution strategy encoding based on attributes is used as the label, and the model is trained through an independent threshold loss function to generate the information distribution strategy.

[0004] However, the above and similar technical solutions still have the following shortcomings in practical applications: simply concatenating multi-source heterogeneous data for classification does not effectively utilize the temporal dependencies in the data, which can easily lead to the drift of feature semantics in different time periods and regions; multi-label classification networks are trained using loss functions with independent thresholds, which can easily lead to low-frequency, high-priority information being incorrectly classified into lower levels in scenarios where multi-dimensional label combinations have a long-tail distribution; the distribution strategy relies on static constraint thresholds, which are difficult to dynamically adjust according to real-time traffic conditions and user cognitive load, resulting in oscillations and convergence difficulties near the decision boundary. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for information classification and intelligent distribution for travel information platforms, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for information classification and intelligent distribution for travel information platforms, comprising: acquiring multi-source heterogeneous time-series data, inputting it into a multimodal feature embedding network, wherein the multimodal feature embedding network dynamically adjusts the historical contribution of the time-series features, and outputs fused features after decoupling the time-series features and spatial features by mutual information minimization constraints, inputting the fused features into a multi-label classification neural network, wherein the parameters of the multi-label classification neural network are maintained by gradient orthogonal projection and branch momentum to obtain a classification result, and inputting the classification result as a state vector into a policy network, wherein the policy network dynamically adjusts the safety boundary of the action based on the safety confidence evaluation of the candidate action according to the current state, and outputs a multi-dimensional action instruction, and pushes information to the target user terminal according to the instruction.

[0007] Furthermore, the multimodal feature embedding network includes parallel temporal and spatial branches; the dynamic adjustment of the historical contribution of the temporal features includes: obtaining the initial temporal feature vector and the historical state representation vector output by the temporal branch at the current time; calculating the distribution difference metric between the initial temporal feature vector and the historical state representation vector; mapping the distribution difference metric to a gating value; using the gating value to perform element-wise weighting on the historical state representation vector; and fusing the weighted historical state representation vector with the initial temporal feature vector to obtain the temporal feature vector at the current time.

[0008] Furthermore, the step of decoupling temporal and spatial features by minimizing mutual information and then outputting fused features includes: obtaining the temporal feature vector output by the temporal branch and the spatial feature vector output by the spatial branch, performing kernel function mapping on each, calculating the mutual information estimate, adding the mutual information estimate as a regularization term and the temporal feature reconstruction loss as an auxiliary constraint to the loss function of the multimodal feature embedding network, updating the feature extraction layer parameters with the goal of minimizing the overall loss, freezing the spatial branch parameters and removing the reconstruction loss when the decoupling degree meets the preset conditions, and concatenating the feature vectors output by the temporal branch and the spatial branch into fused features.

[0009] Furthermore, the gradient orthogonal projection includes: obtaining the gradient vectors of the multi-label classification neural network parameters corresponding to each label category in the current training batch; designating categories with sample frequencies below a preset threshold as tail categories and categories with sample frequencies above a preset threshold as high-frequency categories; calculating the subspace spanned by the gradient vectors of all high-frequency categories; orthogonally projecting the original gradient vectors of the tail categories onto the complement space of the subspace; updating the multi-label classification neural network parameters corresponding to the tail categories using the projected gradient vectors; and updating the high-frequency category parameters while maintaining the original gradient direction.

[0010] Furthermore, the rules for maintaining the split momentum include: maintaining an independent gradient momentum cache for each label category of the multi-label classification neural network; determining the momentum decay coefficient of each label category based on the frequency of occurrence of the corresponding label category in the training samples; having a higher momentum decay coefficient for tail categories than for high-frequency categories; using the momentum cache corresponding to each label category to smooth the parameter gradients of each label category in the current training batch; and updating the parameters of the multi-label classification neural network corresponding to each label category.

[0011] Furthermore, the security confidence score is generated by a confidence evaluation head added after the policy network output layer, and the security confidence score output by the confidence evaluation head represents the probability that the current action satisfies the security constraints.

[0012] Furthermore, the dynamic adjustment of the action safety boundary includes: obtaining the candidate actions output by the policy network in the current state, obtaining the safety confidence score value output by the confidence evaluation head for each candidate action, comparing the safety confidence score value of each candidate action with a preset safety threshold, reducing the safety action boundary of actions below the safety threshold according to the deviation ratio, and widening the safety action boundary of actions above the safety threshold according to the excess range.

[0013] Furthermore, it also includes: classifying actions that violate safety constraints after execution and whose confidence scores are higher than the safety threshold as high-confidence errors, and imposing a penalty proportional to the safety confidence score.

[0014] Furthermore, it also includes: predicting the safety confidence score of the candidate action after execution at a future time based on the current state and the current candidate action; if the prediction result is lower than a preset safety threshold, then reducing the execution magnitude of the current candidate action according to the remaining time steps.

[0015] An information classification and intelligent distribution system for travel information platforms uses any of the aforementioned information classification and intelligent distribution methods for travel information platforms.

[0016] Compared with the prior art, the beneficial effects of the present invention are: A method and system for information classification and intelligent distribution for travel information platforms is proposed. By using a dynamic adjustment mechanism of historical contribution in time-series branches, the gate value directly responds to the degree of semantic drift. Combined with mutual information minimization constraints and reconstruction loss, the method solves the feature semantic drift problem and improves the accuracy of multi-dimensional classification.

[0017] Meanwhile, by orthogonally projecting the gradient of the tail category to the direction decoupled from the high-frequency category, and independently maintaining a frequency-adaptive momentum cache for each category, the gradient suppression problem under the long-tailed distribution of labels is solved, ensuring that low-frequency, high-priority information is correctly classified into higher levels. By using a confidence evaluation head, the safety boundary is endogenously driven by the confidence of the policy network itself. Combined with high-confidence error penalty and trajectory prediction predictive intervention, the decision boundary oscillation problem caused by static constraint threshold is solved, achieving stable adaptive distribution in dynamic environments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the multimodal feature embedding network structure of the present invention; Figure 3 This is a schematic diagram of the multi-label classification neural network parameter update mechanism of the present invention; Figure 4 This is a schematic diagram of the strategy network and security boundary adjustment layer structure of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, the present invention provides a technical solution: a method for information classification and intelligent distribution for travel information platforms, comprising: Acquire multi-source heterogeneous time-series data.

[0021] It is important to note that travel information platforms acquire multi-source heterogeneous time-series data through in-vehicle terminals or mobile terminals, including user location time-series data, motion state time-series data, and interaction behavior time-series data. Location time-series data records the user's geographic coordinate sequence at a fixed sampling frequency (1Hz). Motion state time-series data includes instantaneous velocity, acceleration, and orientation angle, sharing the same timestamps as location data. Interaction behavior time-series data includes user actions such as clicks, browsing, pauses, and returning, recorded in event logs with timestamp precision at the millisecond level. These three types of data are aligned using timestamps as the primary key to form a multi-dimensional time-series matrix, which serves as the input for subsequent multimodal feature embedding networks.

[0022] The multi-source heterogeneous time-series data is input into a multimodal feature embedding network. The multimodal feature embedding network dynamically adjusts the historical contribution of the time-series features and outputs fused features after decoupling the time-series features and spatial features by minimizing mutual information constraints.

[0023] It should be noted that the multimodal feature embedding network includes parallel temporal and spatial branches. Figure 2 This is a schematic diagram of the multimodal feature embedding network structure of this invention. The temporal branch uses a gated recurrent unit (GRU) as the basic encoder, and the hidden layer dimension of the GRU is set to 256 dimensions. At each time t, the temporal branch receives a segment of temporal data within the current input time window. The length of this time window is L, and L is 9. The raw data at each time point within this time window includes GPS coordinates (2D, longitude and latitude), instantaneous velocity (1D), acceleration (1D), and orientation angle (1D). The above 5-dimensional data from the current time t and the four times before and after it (i.e., from t-4 to t+4) are concatenated along the feature dimensions to form a 45-dimensional raw data vector. This vector serves as the input to the GRU encoder at the current time step. If the current time step is located at the edge of the sequence (t < 4 or t > total sequence length - 4), it is filled with the values ​​of the boundary time steps. That is, at the beginning of the sequence, missing preceding time steps are filled with the values ​​of time step t = 0; at the end of the sequence (t > total sequence length - 4), missing subsequent time steps are filled with the values ​​of time step t = total sequence length - 1. The GRU encoder then uses the current input... Perform forward computation with the hidden state from the previous time step, and output the hidden state vector u at the current time step. t Dimension 256. This u t This is the initial temporal feature vector. The historical state representation vector is denoted as H. t-1 This is historical information accumulated from previous time steps, with a dimension of 256, and is updated using a sliding weighted average method.

[0024] The specific process of dynamically adjusting the historical contribution of time series features is as follows: obtain the initial time series feature vector and the historical state representation vector output by the time series branch at the current time; calculate the distribution difference metric between the initial time series feature vector and the historical state representation vector; map the distribution difference metric to a gating value; use the gating value to perform element-wise weighting on the historical state representation vector; and fuse the weighted historical state representation vector with the initial time series feature vector to obtain the time series feature vector at the current time.

[0025] It is important to note that obtaining the initial temporal feature vector u output by the GRU encoder at the current time t is crucial. t and the historical state representation vector H t-1 Using cosine distance as a measure of distributional difference, u is calculated. tWith H t-1 cosine distance value d t : ; d t The value range of d is [0,2]. t The closer d is to 0, the smaller the difference in distribution between the current input and the historical representation. t The closer the value is to 2, the greater the distributional difference. Let d... t The input is fed into a gated mapping module consisting of a single-layer fully connected network and a sigmoid activation function. This fully connected network has an input dimension of 1 and an output dimension of 1. Its weight parameter is w, and its bias parameter is b. The initial values ​​of w and b are -1 and 0.5, respectively, and they are automatically updated during training via backpropagation. The gate value g... t for: ; When the distribution difference d t When it increases, w×d t The value of +b decreases, and g is mapped by the sigmoid function. t Approaching 0, thus inhibiting the transmission of historical information; when d t When it approaches 0, w×d t As the value of +b increases, g t As the value approaches 1, historical information is fully preserved. (Gate value g) t This is the historical contribution adjustment coefficient, with a value range of [0,1]. Using the gate value g... t Historical state representation vector H t-1 By performing element-wise weighting, we obtain the weighted historical state representation vector H. t : ; The weighted historical state representation vector H t With the initial time series feature vector u t By concatenating and fusing the data, we obtain the temporal feature vector h at the current time step. t =[H t ;u t The dimension is 512. It should be noted that the historical state representation vector H... t The update uses the initial time-series feature u t Perform a moving average weighted average, i.e.: ; Here, 0.9 and 0.1 are the preservation coefficient for historical information and the update coefficient for current information, respectively. The sum of these two coefficients is 1, and both are determined through optimization on the validation set during training. This update does not use the concatenated 512-dimensional vector to ensure that the historical representation always accumulates in a 256-dimensional space, maintaining dimensional stability.

[0026] The spatial branch uses an embedding layer to encode location information, mapping each geographic coordinate point to a 256-dimensional spatial feature vector s. t Location information is encoded using geohashing grids. Latitude and longitude coordinates are converted into grid numbers and then input into the embedding layer. The specific implementation of geohashing grid encoding is as follows: the map coverage area is divided into 100 equal parts along both the latitude and longitude directions. Each grid spans 0.009 degrees longitude and 0.009 degrees latitude (approximately 1 kilometer). The user's current latitude and longitude coordinates are mapped to the corresponding grid row and column numbers. The combination of row and column numbers forms a unique grid number, ranging from 0 to 9999. This grid number serves as the index for the embedding layer, which searches for the corresponding 256-dimensional vector as the spatial feature output. The parameter matrix of the embedding layer has a dimension of N×256, where N is the total number of grids. In this embodiment, the map coverage area is divided into 10,000 grids. It should be noted that the total number of grids can be adaptively adjusted according to the actual map coverage area. As the map coverage area expands, the number of divisions is increased accordingly, while maintaining the spatial resolution of each individual grid.

[0027] The specific process of decoupling temporal and spatial features and generating fused features by minimizing mutual information is as follows: obtain the temporal feature vector output by the temporal branch and the spatial feature vector output by the spatial branch, perform kernel function mapping on them respectively, calculate the mutual information estimate, use the mutual information estimate as a regularization term and the temporal feature reconstruction loss as an auxiliary constraint, and add them together to the loss function of the multimodal feature embedding network. Update the feature extraction layer parameters with the goal of minimizing the overall loss. When the degree of decoupling meets the preset conditions, freeze the spatial branch parameters and remove the reconstruction loss. Concatenate the feature vectors output by the temporal branch and the spatial branch to form the fused features.

[0028] It is important to note that the temporal feature vector h is obtained from the output of the temporal branch. t (512-dimensional) and spatial feature vectors s output by spatial branch t (256 dimensions), respectively for h t and s t Perform kernel function mapping. A Gaussian radial basis function kernel is used for mapping, and the kernel function K(a,b) has the following form: ; Where a and b are two vectors whose similarity is to be calculated; ||ab|| is the Euclidean distance between the two vectors; σ is the bandwidth parameter of the kernel function, which takes the value h. t and s t The median of the Euclidean distance between each sample and its k-th nearest neighbor in the feature space, where k is 10. This median is calculated by randomly sampling 1000 samples before training begins. This method allows the kernel function's coverage to adapt to the density of the data distribution, thus ensuring the stability of the mutual information estimation. The Gaussian radial basis function can... t and s t Mapping to a high-dimensional reproducing kernel Hilbert space transforms the originally nonlinearly separable relationship into a linearly separable one, enabling more accurate capture of the nonlinear statistical dependence between two vectors when performing mutual information estimation.

[0029] After performing kernel function mapping, h is computed in the reproducing kernel Hilbert space. t With s t The mutual information estimate MI is calculated using a nonparametric estimation method based on k-nearest neighbors: the k-nearest neighbor distance for each sample is calculated in the joint space, where k is set to 10. The joint entropy and marginal entropy are estimated based on the k-nearest neighbor distances, resulting in the mutual information estimate MI: ; The mutual information estimate MI is added as a regularization term to the overall loss function of the multimodal feature embedding network, and the overall loss function L is: ; Where L1 is the main loss function (cross-entropy loss) for the downstream classification task; λ1 is the weight coefficient of the regularization term; L2 is the reconstruction loss, used to force the temporal branch to retain sufficient original temporal information, prevent excessive compression of features due to mutual information minimization constraints leading to information loss, and ensure the integrity of feature representation. The specific implementation is as follows: a lightweight decoder is added to the output of the temporal branch. This decoder is a two-layer fully connected network. The first layer converts the temporal feature vector h... t The 512-dimensional array is mapped to 256 dimensions, with ReLU as the activation function. The second layer maps the 256 dimensions back to the original input time series data dimensions, outputting a reconstructed time series data vector. Therefore, L2 is: ; λ2 is the reconstruction loss weight coefficient. The weight coefficients λ1 and λ2 are determined by grid search on the validation set. The search range of λ1 is 0.001 to 0.1, and the search range of λ2 is 0.01 to 1.0. The combination that maximizes the classification accuracy of the validation set and results in a mutual information estimate of less than 0.1 is selected. In this embodiment, the search verification shows that λ1=0.01 and λ2=0.1 are the optimal values.

[0030] For the feature extraction layer parameters (including GRU encoder parameters, gating mapping module parameters, and spatial branch embedding layer parameters) of the temporal and spatial branches, backpropagation is performed with the goal of minimizing the overall loss L. Since MI has been added as a regularization term in L, minimizing L naturally makes MI tend to decrease, thereby forcing temporal features and spatial features to become statistically independent, fundamentally cutting off the interference of spatial information on temporal features at the statistical correlation level.

[0031] When the mutual information estimate (MI) continuously decreases to below the absolute threshold of 0.05, or when the MI value decreases by less than 0.5% over 10 consecutive training epochs (whichever occurs earlier), the decoupling degree is determined to meet the preset conditions. The parameters of the spatial branch are frozen (i.e., gradient backpropagation of the spatial branch is stopped, and its parameter values ​​are fixed), and only the temporal branch continues to participate in subsequent training. It should be noted that the absolute threshold of 0.05 and the relative threshold of 0.5% mentioned above are preferred values ​​in this embodiment. In practical applications, these values ​​can be adjusted based on the balance between classification accuracy and mutual information value on the validation set. For example, when the classification accuracy on the validation set is not yet saturated but the mutual information estimate is already below 0.05, the absolute threshold can be appropriately reduced to avoid premature freezing and insufficient information decoupling.

[0032] After freezing the spatial branch, the reconstruction loss L2 and its corresponding lightweight decoder are removed and no longer participate in subsequent loss calculations. The temporal feature vector h output by the temporal branch is... t (Dimension 512) and the spatial feature vector s output by the spatial branch t (Dimension 256) Concatenate along the feature dimension to obtain the fused feature f. t =[h t ;s t The spatial feature vector s is 768 in dimension. After freezing the parameters of the spatial branch, the spatial feature vector s output by the spatial branch is... t Keeping unchanged, fusion feature f t The update only comes from the timing branch h t As training progresses, the spatial branch no longer contributes gradients.

[0033] In this step, a gating mechanism is used to make the forgetting gate behavior of the temporal branch directly respond to the degree of semantic drift in the scene, unlike the conventional GRU where the gating value only depends on a linear combination of the current input and the output of the previous time step. Mutual information minimization constraints eliminate the statistical correlation contamination of temporal features by spatial information. Reconstruction loss prevents the loss of original information in temporal features during decoupling. The synergy of these three mechanisms ensures that the representation vectors of different time periods and regions remain stable, reducing semantic drift.

[0034] The fused features are input into a multi-label classification neural network. The parameters of the multi-label classification neural network are maintained by gradient orthogonal projection and branch momentum to obtain the classification result.

[0035] It is important to note that the multi-label classification neural network consists of a fully connected classification layer. The input dimension of this layer is consistent with the fusion feature dimension, and the output dimension is C, where C represents the total number of categories for the multi-dimensional labels. In this embodiment, C=72, corresponding to all combinations of 4 security level categories, 3 content category categories, 2 time attribute categories, and 3 spatial range categories. The 4 security level categories include: general information, important information, emergency information, and high-risk information, used to control the access permissions and display intensity of information push notifications. The 3 content category categories include: traffic conditions, services, and alerts, used to determine the attribute affiliation of information for personalized filtering. The 2 time attribute categories include: real-time and planned, used to control the timing and validity period of information push notifications. The 3 spatial range categories include: point, line, and area, used to determine the geographical coverage radius of information push notifications. The weight matrix of the fully connected classification layer is denoted as W, with a dimension of 72×768, and the bias vector is denoted as B, with a dimension of 72.

[0036] Figure 3 This is a schematic diagram of the parameter update mechanism of the multi-label classification neural network of the present invention.

[0037] Specifically: The specific process of gradient orthogonal projection is as follows: obtain the gradient vector of the multi-label classification neural network parameters corresponding to each label category in the current training batch, take the category with sample frequency below a preset threshold as the tail category, take the category with sample frequency above the preset threshold as the high-frequency category, calculate the subspace spanned by the gradient vectors of all high-frequency categories, orthogonally project the original gradient vector of the tail category to the complement space of the subspace, update the multi-label classification neural network parameters corresponding to the tail category using the projected gradient vector, and update the high-frequency category parameters while maintaining the original gradient direction.

[0038] It is important to note that for each category c (c=1,2,...,72), the loss function is used to calculate the classifier weight W for that category. c gradient g c The preset threshold τ is determined by the category distribution of the training set. Specifically, the frequency of each category in the training set is counted, sorted from low to high frequency, and the frequency corresponding to the 10th percentile is taken as τ. In this embodiment, τ=50, that is, categories with a frequency below 50 are tail categories, and categories with a frequency above 50 are high-frequency categories.

[0039] Let the number of high-frequency categories be M. Construct a matrix G from the gradient vectors of these M high-frequency categories. high Calculate G highThe subspace S spanned by the column space is denoted as Sorthogonal complement. perp For each tail category c tail The original gradient vector g tail Calculate its position in the orthogonal complement space S perp Projection p on: ; Among them, G high T For G high The transpose of the vector is used. The projected gradient vector p retains only the components orthogonal to the gradient directions of all high-frequency classes. The classifier weights W corresponding to the tail classes are then updated using p. tail The parameters for high-frequency categories are updated while maintaining the original gradient direction: for high-frequency categories, the updated classifier weights... : ; Where η is the learning rate, and η takes a value of 0.001; For the tail category, the updated classifier weights : ; The rules for maintaining the split momentum include: maintaining an independent gradient momentum cache for each label category of the multi-label classification neural network; determining the momentum decay coefficient of each label category based on the frequency of occurrence of the corresponding label category in the training samples; having a higher momentum decay coefficient for tail categories than for high-frequency categories; using the momentum cache corresponding to each label category to smooth the parameter gradients of each label category in the current training batch; and updating the parameters of the multi-label classification neural network corresponding to each label category.

[0040] It is important to note that for each label category c in a multi-label classification neural network, a separate gradient momentum buffer m should be maintained. c The initial values ​​are all set to zero vectors with a dimension of 768. In each training batch, for each class c, the original gradient vector g calculated for the current batch is obtained. c Based on the frequency f of category c in the training samples. c Determine the momentum decay coefficient β for this category. c : ; Where β0 is the basic momentum decay coefficient, with a value of 0.9; γ is the adjustment coefficient, with a value of 0.05; f max This represents the maximum frequency across all categories. If the calculated value on the right exceeds 0.99, then β will be... cTrim to 0.99 to ensure the momentum decay coefficient remains within the valid range of (0,1); if the calculated value on the right is lower than 0.9, keep it unchanged at 0.9 to ensure that the high-frequency category does not experience violent oscillations due to excessively low momentum decay. c The smaller the category, the more log(f) max / f c The larger the value of ), the better. c The larger the value is than β0, the higher the momentum decay coefficient of the tail category is compared to the high-frequency category. Tail categories correspond to higher momentum decay coefficients, meaning their historical gradients have a greater weight in momentum updates, thus maintaining strong update inertia even under sparse sample conditions.

[0041] Update momentum cache: ; in, This is the updated momentum cache; η is the learning rate, with a value of 0.001.

[0042] After smoothing the parameter gradients of the current batch using the updated momentum cache, the classifier parameters for the corresponding categories are updated. ; in, These are the updated classifier weights. The aforementioned branch momentum maintenance and gradient orthogonal projection together constitute the parameter update mechanism of the multi-label classification neural network: in each batch, gradient orthogonal projection is first performed to obtain the gradient p after projection of each tail category, and then the original gradient is replaced by the projected gradient to perform the branch momentum maintenance process.

[0043] The output of a multi-label classification neural network is a 72-dimensional probability vector P, where the c-th dimension P is the probability vector. c y represents the probability that the input sample belongs to category c. The category corresponding to the maximum value of the output vector is taken as the classification result for that dimension. The combination of the classification results of all 72 categories is the final multi-dimensional classification result, denoted as the classification result vector y.

[0044] In this step, gradient orthogonal projection decouples the tail category from the high-frequency category in the update direction, and shunt momentum enhances the continuous optimization capability of the tail category in terms of update inertia. The two work together to ensure that low-frequency, high-priority information can be correctly classified into high-level categories and avoid being suppressed by high-frequency categories.

[0045] The classification result is input as a state vector into the policy network. The policy network dynamically adjusts the safety boundary of the action based on the safety confidence evaluation of the candidate action according to the current state, and outputs a multi-dimensional action instruction. Figure 4 This is a schematic diagram of the strategy network and security boundary adjustment layer structure of the present invention.

[0046] It is important to note that the policy network uses the proximal policy optimization algorithm as the basic reinforcement learning framework. The main structure of the policy network is a three-layer fully connected neural network. The output dimension is consistent with the dimension of the classification result vector y. The dimensions of the two hidden layers are 256 and 128, respectively, and the activation function is ReLU. The output layer dimension corresponds to the total number of discrete action combinations in the action space. The output layer uses the Softmax activation function to output the action probability distribution. The action space is composed of the following discrete parameter combinations: push time is discretized into 8 time periods, including early morning, morning peak, morning, noon, afternoon, evening peak, and night; push priority is divided into 3 levels, including low priority, medium priority, and high priority; content templates are divided into 5 categories, including traffic condition report templates, accident warning templates, construction detour templates, service facility recommendation templates, and comprehensive reminder templates; push frequency is divided into 3 levels, including low frequency (no more than once per hour), medium frequency (2 to 3 times per hour), and high frequency (4 to 6 times per hour); push radius is divided into 3 levels, including short distance (within 500 meters), medium distance (500 to 2000 meters), and long distance (2000 to 5000 meters). The total number of combinations of the above discrete parameters is 1080. The output layer of the policy network is set with 1080 neurons, and each neuron corresponds to a specific combination (time period, priority, template, frequency, radius).

[0047] A confidence evaluation head is added after the action output layer of the policy network. This confidence evaluation head is an independent two-layer fully connected network. Its input is the output of the last hidden layer of the policy network. The intermediate layers are 64-dimensional, using ReLU activation and followed by a dropout layer with a dropout rate of 0.2. The output layer has a dimension of 1080, and after Sigmoid activation, it outputs a confidence vector, where the score q in the i-th dimension is... i This represents the probability that all safety constraints are satisfied after executing the i-th candidate action in the current state. It should be noted that the confidence evaluation head outputs a 1080-dimensional vector during both the training and inference phases to evaluate the safety of all candidate actions at once.

[0048] The security constraints are predefined by the travel information platform and include, but are not limited to: push frequency not exceeding the user-defined upper limit threshold, the security level of the push information not lower than the minimum security level requirement corresponding to the user's current scenario, and push time not falling within the user-defined do-not-disturb period. To verify whether the above constraints are met, a constraint verification rule set is set. This rule set is independent of the network parameters of the policy network. Taking candidate actions as input, it sequentially verifies whether the action meets the preset conditions in the three dimensions of push frequency, security level, and push time. When all three conditions are met, a compliance flag is output; when any condition is not met, a violation flag is output along with the violation type, including frequency exceeding the limit, insufficient level, or time conflict. During the training phase, each experience trajectory generated by the interaction between the policy network and the environment is accompanied by a compliance or violation flag output by this verification rule set. This flag is used to label the training samples of the confidence evaluation head, with a compliance sample label of 1 and a violation sample label of 0, and is also used to determine whether a high-confidence error penalty is triggered. During the inference phase, the compliance or violation flag output by this verification rule set serves as the basis for determining the security boundary, excluding action values ​​that may trigger violations from the safe range. This set of validation rules is only used for data labeling and range filtering during training and inference, and is not used for gradient calculation and parameter updates in the network.

[0049] Training the confidence assessment head. Training data comes from experience trajectories generated during the interaction between the policy network and the environment. Each training sample includes the current state vector, the action sampled by the policy network in that state, and the compliance or violation label output by the verification rule set. An experience replay pool with a capacity of 10,000 samples is set up, and the weight of violation samples is set to 3 times that of compliance samples during sampling. The loss function of the confidence assessment head is a binary cross-entropy loss, which is weighted and superimposed with the policy loss and then backpropagated together. The loss weight coefficient of the confidence assessment head is 0.2. 256 samples are sampled in each batch for training. Training is divided into two phases: warm-up and formal training. In the warm-up phase, the first 1000 trajectories are used only to fill the replay pool and the confidence evaluation head is not updated. After the replay pool is filled, the formal training phase begins. In each gradient update step of the policy network, the confidence evaluation head performs 4 independent gradient updates using the samples in the current replay pool (this update ratio is the preferred value in this embodiment, and can be adjusted according to the rate of change of the calibration error of the confidence evaluation head on the validation set in actual applications). After the replay pool is full, the old samples are replaced in a first-in-first-out manner.

[0050] The specific process of dynamically adjusting the safety boundary of an action is as follows: obtain the candidate actions output by the policy network in the current state, obtain the safety confidence score value output by the confidence evaluation head for each candidate action, compare the safety confidence score value of each candidate action with a preset safety threshold, reduce the safety action boundary of actions that are lower than the safety threshold according to the deviation ratio, and widen the safety action boundary of actions that are higher than the safety threshold according to the excess range.

[0051] It's important to note that in the current state (i.e., the classification result vector y), the policy network's main output layer outputs a 1080-dimensional original action probability distribution. The policy network then samples categories based on this probability distribution to obtain candidate actions, which are 1080-dimensional one-hot encoded vectors. The confidence evaluation head calculates the safety confidence score corresponding to each candidate action in the current state. After obtaining the safety confidence score of each candidate action, the safety boundary adjustment layer first determines the safety boundary, and then samples actions within this boundary. Specifically, the confidence score q of each candidate action is calculated... i With the preset safety threshold q t The preset security threshold q is compared. t The determination method is as follows: On the validation set, traverse the thresholds from 0.5 to 0.95 with a step size of 0.05. Calculate the comprehensive reward value for each candidate threshold, and select the threshold that maximizes the comprehensive reward value. The calculation method for the comprehensive reward value is as follows: ; Among them, the safety constraint satisfaction rate is the proportion of actions that satisfy all safety constraints on the validation set, and the average task reward is R on the validation set. task The arithmetic mean of (+1 for user click, -1 for close, 0 for no operation). The security threshold q in this embodiment. t Take 0.7.

[0052] The safety boundary adjustment layer defines the safety boundary as a constraint box containing three subsets of action parameters (adjustable range of push time, adjustable range of push priority, and adjustable range of push frequency). For the safety confidence score q... i Below q t The action dimension, and the adjustable range of each of the three dimensions are multiplied by a scaling factor θ1: ; For the safety confidence score q i Higher than q t The action dimension, and the adjustable range of each of the three dimensions are multiplied by the expansion coefficient θ2: ; The maximum relaxation is 1.3 times the original range. The safety boundary adjustment layer prioritizes pruning to handle sampling results that exceed the boundary. If more than three dimensions exceed the boundary simultaneously in a single sampling, the sampling result is discarded, and resampling is performed within the boundary range (probabilities outside the boundary are set to zero, and probabilities inside the boundary are renormalized). The action output by the safety boundary adjustment layer is the final execution instruction, which is a 1080-dimensional one-hot encoded vector.

[0053] In this step, the safety boundary is intrinsically driven by the policy network's own evaluation of action safety, rather than relying on externally preset static thresholds. This eliminates the conflict between external constraints and the policy network output, and suppresses oscillations near the decision boundary.

[0054] During the training phase, actions that violate safety constraints after execution and whose confidence scores are higher than the safety threshold are judged as high-confidence errors and are subject to penalties proportional to the safety confidence scores.

[0055] It is important to note that during the training phase, after the policy network interacts with the environment, the actual security effect after each action is executed is recorded. If the security confidence score of a certain action is q... i Higher than q t However, if the action violates safety constraints after execution, it is classified as a high-confidence error, and an additional penalty term proportional to the safety confidence score is added to the overall reward function of the near-end policy optimization algorithm. The overall reward function R is of the following form: ; Among them, R task The system provides rewards for user feedback (+1 for a click, -1 for closing, 0 for no action); α is the penalty coefficient, which is selected by iterating through the validation set at steps of 0.5, ranging from 0.5 to 5.0, to ensure the safety constraint satisfaction rate is higher than 95% and the average task reward is the highest. In this embodiment, α is set to 2.0; I is an indicator variable, set to 1 when the action violates the safety constraint, and 0 otherwise. This penalty mechanism guides the policy network to avoid overconfident outputs on high-risk actions during the learning process.

[0056] During the inference phase, the system also includes predicting the safety confidence score of the candidate action at a future time based on the current state and the current candidate action. If the prediction result is lower than a preset safety threshold, the execution magnitude of the current candidate action is reduced proportionally to the remaining time steps.

[0057] It is important to note that during the inference phase, the policy network no longer updates its parameters, only performing forward inference. Before executing an action, the safety boundary adjustment layer adds a short-term trajectory prediction step: obtaining the current state and the current candidate action. The current state is the classification result vector y, output in real-time by a multi-label classification neural network from the current input sample; the current candidate action is the original action sampled by the main output layer of the policy network before safety boundary adjustment, i.e., a 1080-dimensional one-hot encoded vector that has not yet undergone boundary clipping or resampling. The current state and the current candidate action are input into a pre-trained state transition prediction network, a three-layer fully connected network with an input dimension of 72 (state dimension) plus an action dimension of 1080, two hidden layer dimensions of 256 and 128 respectively, and an output dimension of 72, predicting the state estimate for the next time step after the action is executed.

[0058] The training of the state transition prediction network is performed synchronously with the policy network training phase, sharing the same trajectory data in the experience replay pool as the confidence evaluation head. After each round of policy network interaction, a batch of continuous trajectory segments is randomly sampled from the replay pool, each segment being 8 time steps in length. The current state and the action performed at each time step are used as input, and the observed state at the next time step is used as the supervision signal to minimize the mean square error between the predicted state and the true state. The state transition prediction network is updated once per round, consistent with the update frequency of the policy network. In the initial stage of training, the first 500 trajectories are only used to fill the replay pool and are not used to update the parameters of the state transition prediction network; after the replay pool is full, the state transition prediction network begins formal training. During the inference phase, the network is no longer updated and only performs forward inference.

[0059] Using a trained state transition prediction network, starting from the current state, and with the current candidate action fixed, iteratively predict the state sequence for the next T time steps, where T is 5. Each predicted future state is input into the confidence evaluation head to obtain the corresponding sequence of predicted safety confidence scores. If the q value at any time in the prediction sequence is lower than the preset safety threshold (0.7), the safety boundary adjustment layer reduces the execution amplitude of the current candidate action proportionally to the remaining time steps. In the multidimensional discrete action space, the action execution amplitude refers to the degree of deviation of the values ​​of each dimension of the action. Specifically, each discrete action dimension has a default value (the push period defaults to the current time period, the priority defaults to medium, and the push frequency defaults to medium). The execution amplitude is defined as the adjustment amplitude of each dimension's value relative to the default value. When a buffer coefficient ρ is applied, the discrete values ​​of each dimension output by the policy network are retained with probability ρ and regressed to the default value with probability (1-ρ). Specifically, let the first prediction time below 0.7 be t+j (1≤j≤T), then the execution amplitude of the current candidate action is multiplied by the buffer coefficient ρ, and the formula for calculating ρ is: ; For example, if the prediction is to enter the low-confidence region at the third future time (j=3, T=5), then ρ=2 / 5=0.4, and the current action's magnitude is reduced to 40% of its original value. If the q value is not lower than 0.7 at all times in the prediction sequence, then ρ=1, and the action's magnitude remains unchanged. This mechanism allows the policy network to gradually reduce the action magnitude before entering the low-confidence region, avoiding sudden changes at the boundary that could cause decision oscillations.

[0060] The information is pushed to the target user terminal according to the instructions.

[0061] It is important to note that the multi-dimensional action instructions ultimately output by the policy network include the push period, push priority level, content template number, push frequency level, and push radius level. The travel information platform's distribution module receives these instructions, determines the sending time according to the push period, determines the order in the message queue according to the push priority level, selects the corresponding information content template according to the content template number and fills it with the current travel information data, determines the minimum time interval between two adjacent pushes according to the push frequency level, and determines the geographical coverage area of ​​the information according to the push radius level. The distribution module then sends the assembled information content to the target user terminal device through the push gateway. User terminal devices include, but are not limited to, smartphones, in-vehicle terminals, and smart wearable devices.

[0062] An information classification and intelligent distribution system for travel information platforms utilizes any of the aforementioned information classification and intelligent distribution methods for travel information platforms. The system includes: a data acquisition module for acquiring multi-source heterogeneous time-series data; a multimodal feature embedding module for dynamically adjusting the historical contribution of time-series features and decoupling time-series and spatial features using mutual information minimization constraints to output fused features; a multi-label classification module for classifying the fused features and outputting classification results; a strategy distribution module for generating multi-dimensional action instructions based on the classification results; and a push execution module for pushing information to target user terminals based on the multi-dimensional action instructions. These modules are deployed on the cloud server of the travel information platform and perform forward computation by calling a machine learning inference engine.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for information classification and intelligent distribution for travel information platforms, characterized in that, include: Multi-source heterogeneous time-series data is acquired and input into a multimodal feature embedding network. The multimodal feature embedding network dynamically adjusts the historical contribution of the time-series features and decouples the time-series features and spatial features by minimizing mutual information, outputting fused features. The fused features are then input into a multi-label classification neural network. The parameters of the multi-label classification neural network are maintained by gradient orthogonal projection and branch momentum to obtain the classification result. The classification result is used as a state vector and input into a policy network. The policy network dynamically adjusts the safety boundary of the action based on the safety confidence evaluation of the candidate action according to the current state and outputs a multi-dimensional action instruction. Based on the instruction, information is pushed to the target user terminal.

2. The information classification and intelligent distribution method for a travel information platform according to claim 1, characterized in that: The multimodal feature embedding network includes parallel temporal and spatial branches; The dynamic adjustment of the historical contribution of the time series features includes: obtaining the initial time series feature vector and the historical state representation vector output by the time series branch at the current time; calculating the distribution difference metric between the initial time series feature vector and the historical state representation vector; mapping the distribution difference metric to a gating value; using the gating value to perform element-wise weighting on the historical state representation vector; and fusing the weighted historical state representation vector with the initial time series feature vector to obtain the time series feature vector at the current time.

3. The information classification and intelligent distribution method for a travel information platform according to claim 1, characterized in that: The process of decoupling temporal and spatial features by minimizing mutual information and then outputting fused features includes: obtaining the temporal feature vector output by the temporal branch and the spatial feature vector output by the spatial branch, performing kernel function mapping on each, calculating the mutual information estimate, adding the mutual information estimate as a regularization term and the temporal feature reconstruction loss as an auxiliary constraint to the loss function of the multimodal feature embedding network, updating the feature extraction layer parameters with the goal of minimizing the overall loss, freezing the spatial branch parameters and removing the reconstruction loss when the decoupling degree meets the preset conditions, and concatenating the feature vectors output by the temporal and spatial branches into fused features.

4. The information classification and intelligent distribution method for a travel information platform according to claim 1, characterized in that: The gradient orthogonal projection includes: obtaining the gradient vectors of the multi-label classification neural network parameters corresponding to each label category in the current training batch; designating categories with sample frequencies below a preset threshold as tail categories and categories with sample frequencies above a preset threshold as high-frequency categories; calculating the subspace spanned by the gradient vectors of all high-frequency categories; orthogonally projecting the original gradient vectors of the tail categories onto the complement space of the subspace; updating the multi-label classification neural network parameters corresponding to the tail categories using the projected gradient vectors; and updating the high-frequency category parameters while maintaining the original gradient direction.

5. The information classification and intelligent distribution method for a travel information platform according to claim 1, characterized in that: The rules for maintaining the split momentum include: maintaining an independent gradient momentum cache for each label category of the multi-label classification neural network; determining the momentum decay coefficient of each label category based on the frequency of occurrence of the corresponding label category in the training samples; having a higher momentum decay coefficient for tail categories than for high-frequency categories; using the momentum cache corresponding to each label category to smooth the parameter gradients of each label category in the current training batch; and updating the parameters of the multi-label classification neural network corresponding to each label category.

6. The information classification and intelligent distribution method for a travel information platform according to claim 1, characterized in that: The security confidence score is generated by a confidence evaluation head added after the policy network output layer. The security confidence score output by the confidence evaluation head represents the probability that the current action satisfies the security constraints.

7. The information classification and intelligent distribution method for a travel information platform according to claim 1, characterized in that: The dynamic adjustment of the action safety boundary includes: obtaining the candidate actions output by the policy network in the current state, obtaining the safety confidence score value output by the confidence evaluation head for each candidate action, comparing the safety confidence score value of each candidate action with a preset safety threshold, reducing the safety action boundary of actions below the safety threshold according to the deviation ratio, and widening the safety action boundary of actions above the safety threshold according to the excess range.

8. The information classification and intelligent distribution method for a travel information platform according to claim 7, characterized in that: Also includes: Actions that violate safety constraints after execution and whose confidence scores are higher than the safety threshold are judged as high-confidence errors and are subject to penalties proportional to the safety confidence scores.

9. The information classification and intelligent distribution method for a travel information platform according to claim 7, characterized in that: Also includes: Based on the current state and the current candidate action, predict the safety confidence score after executing the candidate action at a future time. If the prediction result is lower than the preset safety threshold, reduce the execution magnitude of the current candidate action according to the remaining time steps.

10. An information classification and intelligent distribution system for travel information platforms, characterized in that, The method for information classification and intelligent distribution for a travel information platform, as described in any one of claims 1-9, is used.

Images