Advertisement putting method and system based on charging pile

Abstract

The invention relates to the technical field of new energy charging and advertisement putting, and discloses an advertisement putting method and system based on a charging pile, and the method comprises the steps: collecting user interaction, behavior, position, time and environment data, and constructing a behavior feature data set; constructing a third-order behavior tensor based on the behavior characteristics, and decomposing to generate a user behavior embedding vector; calculating an information entropy value of advertisement selection; screening the target advertisement with the maximum entropy reduction amplitude for putting; displaying the target advertisement, collecting user response and updating behavior characteristics; retraining the behavior tensor model to adjust the advertisement strategy; the system comprises a data acquisition module, a tensor modeling module, a strategy optimization module, a strategy screening module, an edge putting module and a strategy updating module. According to the method, the behavior tensor taking the user identifier, the time slice and the geographic position as the three-dimensional index is constructed, and the behavior embedding vector is extracted by introducing the tensor decomposition method, so that the high-dimensional semantic modeling of the user behavior is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of new energy charging and advertisement delivery, in particular to an advertisement delivery method and system based on charging piles. BACKGROUND

[0002] With the rapid development of the new energy automobile industry, the layout density of urban public charging piles is continuously improved, and gradually extends from "infrastructure" to "interactive carrier". In the user charging process, the user's attention is relatively concentrated, the staying time is controllable, and a natural scene for advertisement display is objectively formed. Such a scene has strong physical location attributes, clear time period characteristics and user behavior characteristics highly related to vehicle use habits, and is therefore widely considered to have high information reach potential.

[0003] Existing advertisement delivery technologies are mostly deployed by means of existing mobile advertisement technology frameworks. The basic idea is to use charging equipment as an advertisement terminal, and to statically match the basic identity information (such as account, vehicle type, region, etc.) of the user with preset advertisement tags. The advertisement selection logic often relies on preset rules, targeting strategies or simple classification algorithms for delivery, and the content is mostly a unified advertisement package set based on region, time or service provider background, lacking in-depth analysis of the dynamic behavior characteristics of individual users.

[0004] However, the existing advertisement delivery technology relies too much on static tags for user portrait construction, and is difficult to integrate multi-dimensional data such as location information, charging behavior and environmental variables, resulting in a single and stereotyped portrait. Secondly, the advertisement delivery strategy lacks the ability to dynamically capture the evolution of user behavior, and is difficult to adjust the content distribution path in a timely manner according to user feedback. Thirdly, in the recommendation model training mechanism, the traditional way generally uses a centralized periodic update mode, which has low update frequency and slow response. Therefore, the present application provides an advertisement delivery method and system based on charging piles to solve the problems existing in the prior art. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide an advertisement delivery method and system based on charging piles, which solves the problems of single user portrait dimension, delayed behavior response and slow strategy update in the existing advertisement delivery technology.

[0006] To achieve the above purpose, the present application is implemented by the following technical solution: an advertisement delivery method based on charging piles, comprising the following steps: S1, collecting interaction data of the user and the charging pile, user behavior data, location information, time information and environmental state data, and constructing a data set of behavior characteristics; S2, based on the data set of behavior characteristics, a three-order behavior tensor with user identification, time segment and geographic location as dimensions is constructed, and the three-order behavior tensor is decomposed to generate a corresponding user behavior embedding vector; S3, the user behavior embedding vector is matched with a candidate advertisement set, a probability distribution model of the user for each candidate advertisement is established, and the information entropy value of the advertisement selection is calculated based on the probability distribution model of each candidate advertisement; S4, the target advertisement causing the largest entropy drop amplitude is identified and screened based on the information entropy value, as the current advertising; S5, the target advertisement is displayed on the charging pile terminal, the response behavior data of the user to the target advertisement is collected, and the user behavior characteristics are updated based on the response behavior; S6, the updated user behavior characteristics are fed back to the cloud system for retraining of the behavior tensor model and dynamic adjustment of the advertising strategy.

[0007] Preferably, in step S1, the data set of behavior characteristics includes the following steps: Based on the collected user and charging pile interaction data, a user-time interaction mapping table is constructed, and by matching the location information and weather data, a behavior context with space-time scene semantics is generated; A multi-dimensional data structure is constructed driven by the behavior context, wherein each behavior record is mapped to a structured vector containing user identification, charging duration, time label, weather state and scene number; The structured vector is used as the data basis in the subsequent tensor modeling step and is input into the tensor dimension generator to define the index domain of the behavior tensor.

[0008] Preferably, in step S2, the generation of the corresponding user behavior embedding vector includes the following steps: Based on the generated structured vector, the user number, time segment number and geographic area number are respectively taken as the three dimension identifiers of the three-order behavior tensor; In the tensor construction process, the existence of behavior events is taken as the tensor value setting criterion to form a three-order sparse behavior tensor; The CP decomposition method of minimizing reconstruction error is used to decompose the tensor to generate a user embedding vector.

[0009] Preferably, the CP decomposition method of minimizing reconstruction error is used to decompose the tensor, and the minimizing reconstruction error uses the following formula: ; In the formula, represents the interaction behavior data of user, time and location, represents the rank-one component on the user, time and location dimensions, respectively, dimension of the latent space, denotes the vector outer product.

[0010] Preferably, in step S3, the calculation of the information entropy value of the advertisement selection based on the probability distribution model of each candidate advertisement comprises the following steps: The user embedding vector is input into the matching score network, and inner product operation is performed with each advertisement feature vector to generate a matching score; The selection probability of the user for each advertisement is calculated by Softmax normalization of all advertisement scores, and the Softmax normalization calculation formula is: ; In the formula, denotes the probability of the user selecting the advertisement, denotes the exponential summation of all advertisement scores, the score value of the first advertisement, is the base of the natural logarithm; According to the probability distribution, an information entropy function is constructed , which is used to quantify the distribution of the current advertisement; In the formula, is the information entropy of the advertisement selection, denotes the probability of the user selecting the advertisement, is the logarithm of the selection probability of the first advertisement, is the total number of advertisements.

[0011] Preferably, in step S4, the identification and screening of the target advertisement causing the largest entropy decrease based on the information entropy value comprises the following steps: The established probability distribution model is called to simulate the change of user behavior after each candidate advertisement is launched, and a predicted probability distribution is generated; For each candidate advertisement, the information entropy change is calculated as the information structure change after the advertisement is launched; Based on the information entropy change results of all candidate advertisements, the advertisement with the largest information entropy decrease is screened and set to meet the constraint of information entropy decrease, as the final launched advertisement.

[0012] Preferably, in step S5, the updating of the user behavior feature based on the response behavior comprises the following steps: The display record of the launched advertisement is combined with the user click, browsing time and interaction path to generate a response vector; The update gradient is calculated based on the response vector and the launched advertisement vector, which is used to correct the user factor vector in the tensor decomposition model in reverse; The user factor vector is modified, and the change in the behavior preference of the user in the third-order tensor is dynamically mapped.

[0013] Preferably, the step of calculating the update gradient based on the response vector and the advertising vector includes the following steps: The response vector of the user to the advertising is set as , the embedding vector of the advertising is , and the current user factor vector is ; The loss function for calculating the user behavior preference error is constructed as: ; In the formula, is the loss function, , and the matching function of the predicted user response is represented by The update gradient of the user factor vector is obtained by deriving the calculated loss function with respect to ; In the formula, is the gradient of the loss function with respect to the user factor vector , the partial derivative of the loss function with respect to , the predicted advertising response is , and the derivative of the activation function

[0014] Preferably, in step S6, the step of retraining the behavior tensor model and dynamically adjusting the advertising strategy includes the following steps: The updated user embedding vector changes are summarized, and the feature data of the vector change direction and amplitude are uploaded to the cloud; The cloud system uses the change summaries uploaded by multiple edge nodes to construct a sample set for fine-tuning the tensor model, and re-trains the latent factor parameter set based on the sample set; After the training is completed, the updated parameters are distributed to the edge devices to complete the distributed collaborative evolution of the strategy model.

[0015] An advertising delivery system based on charging piles is also provided, which includes: A data acquisition module for acquiring various types of data, including user behavior data, device state data, scene environment data, and advertising response data; A tensor modeling module for constructing a third-order behavior tensor based on the data provided by the data acquisition module; ​​​​A strategy optimization module is configured to calculate a probability distribution of user acceptance of an advertisement based on the generated user behavior embedding vector, and to generate an optimal delivery strategy by using information entropy to evaluate the potential effect of advertisement acceptance. A strategy screening module is configured to screen out optimal candidate advertisements according to the change amount of information entropy based on the optimal delivery strategy. An edge delivery module is configured to perform actual pushing of the advertisement at the charging pile end and to evaluate the effect of the advertisement by collecting feedback information. A strategy updating module is configured to update the parameters of the delivery strategy after receiving the feedback information of the edge delivery module, and to synchronize the updated strategy to the devices at the charging pile end.

[0016] In summary, the present application includes at least one of the following beneficial technical effects: 1. The present application realizes high-dimensional semantic modeling of user behavior by constructing a behavior tensor indexed by user identification, time segment and geographic location, and introducing a tensor decomposition method to extract behavior embedding vectors. Compared with the traditional advertisement system which only relies on click rate, labeled user features and other static indicators, it is difficult to capture deep preference associations. This method more accurately depicts the user interest evolution path, effectively addressing the technical shortcomings of coarse user profile granularity and single behavior modeling.

[0017] 2. The present application uses information entropy reduction as the basis for selecting target advertisements, and in the advertisement screening process, it no longer simply relies on matching score ranking, but introduces a structural index of user selection uncertainty. This processing method breaks out of the limitations of the linear recommendation framework and can dynamically identify advertisements that truly have "information focusing" capabilities. It solves the problem of low click-through rate and low conversion rate caused by the tendency of previous recommendation systems to push redundant and attention-dispersing advertisement segments.

[0018] 3. The present application introduces a mechanism for updating user factor vectors after advertisement response, and performs gradient back correction with the response vector by constructing a loss function. The user profile has the ability to continuously learn and correct with advertisement interaction. Unlike the periodic full retraining in existing technologies, this solution can achieve rapid adaptation at the edge level, solving the efficiency bottleneck of delayed response in behavior modeling.

[0019] 4. The present application combines the user embedding vector variation summary uploaded by multiple source edge devices, and the system performs tensor model fine-tuning training in the cloud, and completes the synchronization and cooperation of the model in the edge node through parameter distribution. This design discards the high-cost mode of traditional centralized large model global update, and realizes distributed lightweight optimization. Especially in the application environment of changing scenes and real-time user behavior, it avoids the recommendation failure caused by model lag, significantly improving the adaptability and robustness of the system in actual operation. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a method step flowchart of the present application; Figure 2 is a system architecture diagram of the present application. DETAILED DESCRIPTION

[0021] The following will be described in detail below with reference to the accompanying Figure 1 - the accompanying drawings Figure 2 , the present application will be further described in detail.

[0022] Please refer to the accompanying Figure 1 , the present application provides a security person and vehicle multi-modal retrieval engine method, comprising the following steps: S1, collecting user interaction data, user behavior data, location information, time information and environmental state data of the user and the charging pile, and constructing a data set of behavior characteristics; S2, based on the data set of behavior characteristics, constructing a three-order behavior tensor with user identification, time segment and geographic location as dimensions, and decomposing the three-order behavior tensor to generate corresponding user behavior embedding vectors; S3, matching the user behavior embedding vectors with a candidate advertisement set, establishing a probability distribution model of the user for each candidate advertisement, and calculating the information entropy value of the advertisement selection based on the probability distribution model of each candidate advertisement; S4, identifying and screening the target advertisement that causes the largest entropy drop amplitude based on the information entropy value, as the current advertising; S5, displaying the target advertisement on the charging pile terminal, collecting the response behavior data of the user to the target advertisement, and updating the user behavior characteristics based on the response behavior; S6, feeding back the updated user behavior characteristics to the cloud system for retraining the behavior tensor model and dynamically adjusting the advertising strategy.

[0023] For step S1, in the present embodiment, in order to overcome the problems of rough user portrait, single label dimension and the like of the existing advertising system in the charging pile scene, a behavior data driven portrait modeling mechanism is introduced at the charging pile end. By collecting direct interaction information between the user and the device, auxiliary elements such as environmental state, time period and spatial distribution are concerned. A behavior characteristic data set with spatiotemporal semantic, behavior context and user identification fusion capability is formed.

[0024] When the charging behavior occurs, the system first establishes the user-time interaction mapping relationship to construct the original behavior event table. For example, a user U charges at a certain commercial charging station B on March 15, 2025 at 18:45, and the charging start and end time, device ID, power usage and the like will be collected in real time.

[0025] As a possible implementation, the behavior event is associated with the location information database to obtain the coding information L of the geographical area to which the charging pile belongs. At the same time, the system calls an external interface to obtain the weather data corresponding to the time point, including temperature, weather type (sunny / rain / snow, etc.), humidity level, etc., to generate the behavior context field.

[0026] Specifically, the behavior context can be encoded as the following structured entry: User identifier (e.g., U number); Charging start time, end time, duration; Geographic location code (e.g., GPS / neighborhood number); Time label (e.g., weekday / weekend, morning rush hour / night rush hour); Weather state code (e.g., 01 for sunny, 03 for light rain); Scene number (optional, selected from a defined scene dictionary).

[0027] In one possible solution, the above structure is further processed into the index form required by the multi-dimensional tensor input format. A three-dimensional tensor is constructed , where each dimension has the following meanings: represents a set of user numbers, with the number range being ; represents a set of time segments, divided by hours or natural blocks; represents a set of geographic location codes, defined by regions or neighborhoods.

[0028] Each structured behavior vector can be mapped to a non-zero element in the tensor, and its value can be set as a behavior intensity indicator (such as charging duration, interaction frequency, etc.), or a Boolean value to indicate the occurrence of the event. This tensor will be used as the basic data structure for potential factor modeling in subsequent steps.

[0029] As an option, to enhance the timeliness and update ability of the data, the system can set a sliding window mechanism. For example, the user behavior mapping set is refreshed every 24 hours, only the last effective behaviors are retained, avoiding graph expansion and time distortion.

[0030] In actual application scenarios, for example, when users show different charging preferences under different weather conditions (e.g., preferring to use underground parking lot charging in rainy weather), the behavior context can reflect this potential behavior variability and provide an important basis for subsequent advertisement scenario matching.

[0031] In some implementations, the system can integrate additional interfaces with carrier data to collect implicit features such as user communication activity, movement patterns, and app usage frequency, further enriching the profile input space. Subject to data authorization, POS payment records and consumer preference types can also be incorporated as auxiliary tags for multi-dimensional modeling.

[0032] This constructed dataset not only serves as the subsequent tensor structure generation but also serves as input for training deep behavior prediction models. During the model training phase, this dataset is fed into the user behavior preference model in batches, which is optimized by minimizing the difference between predicted and actual behavior.

[0033] For example, the system can introduce the following baseline loss function at the beginning of construction: ; Where, is the total number of data points; For the Behavioral data, is the model prediction value, and the loss function is used to measure the behavioral feature restoration error. The smaller it is, the better the data structure restoration effect is.

[0034] Generally speaking, the system will perform stratification and labeling of users at this stage, dividing them into behavioral clusters such as "highly active", "cyclical", and "weekend consumption" to guide subsequent advertising matching models to adopt different strategies.

[0035] Regarding step S2, in this embodiment, after completing the construction of the user behavior feature data set, it is necessary to further map this multi-source fused structured behavior data into a high-dimensional behavior space to achieve a quantitative expression of individual user behavior preferences. To this end, in this step, the system uses tensor modeling and decomposition methods to extract the latent factor features implicit in the spatiotemporal behavior dimensions, and ultimately generates an embedded representation of user behavior, which serves as the input basis for the subsequent advertising strategy model.

[0036] In this embodiment, it mainly includes: construction of a third-order behavior tensor and generation of a user embedding vector based on CP decomposition.

[0037] Specifically, the system first encodes the structured identification items in the user behavior events: Encode user ID as user dimension index ; Encode time segment information as time dimension index , which can be divided by hour, intraday cycle or behavior block; Mapping geographic area codes to spatial location dimension indexes The GPS grid, administrative region, or commercial area number can be set.

[0038] In some embodiments, whether a behavior event occurs can be marked as a Boolean variable, and the tensor value setting criteria are defined as follows: If the index corresponding to the position has a user behavior event (such as user charging behavior in the area in the time slice , the third-order tensor , otherwise, .

[0039] In general, the tensor has a high degree of sparsity, and only the non-zero items are assigned at the user's real behavior points.

[0040] After the tensor is constructed, the system uses the CP decomposition method to minimize the reconstruction error to model it. This method represents the tensor as the outer product of multiple rank-one tensors, and the goal is to approximate the original behavior tensor as the product of three low-rank embedding matrices. The optimization objective function is as follows: ; In the formula, represents the interaction behavior data of users, time, and location; represents the rank-one component in the user, time, and location dimensions, respectively; represents the vector outer product; the parameter represents the decomposition rank, which is usually a hyperparameter and is determined by actual performance tuning; represents the Frobenius norm, which is used to measure the error of the tensor approximation reconstruction.

[0041] In one possible implementation, the system uses random initialization of latent factor matrices and iteratively solves the minimization problem through methods such as alternating least squares (ALS) or Adam optimizer.

[0042] Finally, the result of tensor decomposition will generate a set of embedding vectors, where each user corresponds to a set of vectors representing their behavior preference features in the latent semantic space. The embedding vector will be one of the inputs to the subsequent ad matching scoring model, and will play a role in characterizing the user's long-term preferences and behavior characteristics.

[0043] In some extended implementations, to improve expressiveness, the system can also embed the tensor modeling process into a neural network structure to build a deep tensor decomposition model (such as TensorFusionNetwork), and simultaneously optimize the embedding parameters and policy network through end-to-end training, thereby enhancing the model's fitting ability for complex behavior patterns.

[0044] Notably, the user embedding vector not only contains the statistical features of the user's historical behavior, but also implies the joint preference for the time and spatial location dimensions. This enables the subsequent strategy model to more accurately identify the user's response tendency in a specific scenario when processing dynamic scenario matching.

[0045] For step S3, in this embodiment, after completing the generation of the user behavior embedding vector, to achieve accurate screening and sorting of personalized advertisements, the system needs to establish a response probability prediction model of the user to the candidate advertisement set based on the embedding vector. Not only is it used to estimate the user's tendency to click or focus on a specific advertisement, but it also provides numerical basis for information entropy evaluation and target advertisement screening strategy.

[0046] An advertisement probability distribution prediction function is constructed, and the uncertainty of advertisement selection is quantified based on the distribution. This process considers the expression ability of the advertisement feature vector and the semantic matching degree between the user behavior embedding vector.

[0047] In general, the system will maintain a set of feature vectors containing multiple advertisement candidates, denoted as , where each vector represents the feature representation of the th candidate advertisement in the dimension semantic space.

[0048] The user's behavior embedding vector is derived from the aforementioned tensor decomposition process. The system inputs to the scoring matching module and performs matching calculation with all candidate advertisement vectors. As an option, the matching function can use the dot product method, i.e.: ; In the formula, is the matching score of the user to the advertisement . The higher the score, the stronger the user's preference for the advertisement. After the matching score calculation is completed, the system inputs the score set to the Softmax function and normalizes it to a probability distribution form to obtain the selection probability of the user for each advertisement. The expression of the Softmax function is as follows: ; In the formula, represents the probability of the user selecting the advertisement; represents the exponential sum of all advertisement scores; the score value of the th advertisement; is the base of the natural logarithm.

[0049] In one possible implementation, the system will calculate the Composition probability distribution ,This distribution not only reflects the user’s advertising response tendency in the current ,scenario, but also indirectly characterizes the distribution density of behavioral ,preferences.

[0050] To measure the degree of certainty in the distribution of ad responses, the system further introduces the information entropy metric to characterize the structural complexity of ad selection. A higher entropy value indicates a more dispersed distribution of user interest in each ad; a lower entropy value indicates a more focused user interest and a clearer ad preference. The information entropy function is defined as follows: ; in, The entropy value representing the probability distribution of current ad selection; is a logarithmic function, usually with natural logarithm as base; User's The probability of selecting an ad comes from the Softmax normalization result; is the total number of ads.

[0051] As an option, to improve the real-time and scalability of the system, probability calculation and entropy evaluation can be performed in parallel, and some model parameters can be preloaded in the edge computing node to achieve lightweight deployment.

[0052] In some embodiments, the system can also consider the impact of the order in which advertisements are displayed, by adding advertisement context factors. right Perform dynamic corrections and build a context-aware matching score function: ; Where, For users to advertise Matching score; To calculate the inner product between the user vector and the ad correction vector; For users The behavioral feature embedding vector.

[0053] The above processing can further improve the model's adaptability to scenarios, and is particularly suitable for charging scenarios where the layout of advertising spaces is complex or the timing of user behavior is sensitive.

[0054] Regarding step S4, in this embodiment, based on the established user-ad matching probability distribution, relying solely on the highest-probability ad for direct delivery can lead to repetitive push notifications or diluted interest. To further improve the matching stability between ad content and user status, an ad screening strategy driven by information entropy changes is introduced.

[0055] By quantifying the structural changes in users' probability of selecting an ad before and after an ad is delivered, the system can dynamically identify candidate ads that effectively reduce user decision-making uncertainty. Ultimately, using the magnitude of entropy reduction as a screening criterion, it identifies the target ads with the greatest structural gain in delivery effectiveness.

[0056] In this embodiment, the system is based on the advertising probability distribution obtained in the previous stage. , first construct its information entropy function to represent the current uncertainty state of the advertising response distribution: ; Where, Information entropy selected for the ad; Indicates the probability that the user selects the ad; For the The logarithm of the probability of selecting an ad; is the total number of ads.

[0057] Then, the system will process each candidate ad. Specifically, the probability distribution change of the user behavior state after the advertisement is simulated is recorded as This distribution can be achieved by introducing a delivery response offset factor. For example, based on the user's past response model, the potential shift in preferences after delivery is predicted.

[0058] In a possible implementation, the system introduces a prediction module to predict the The click rate, dwell time or conversion rate after the The prediction module can be composed of a lightweight neural network, which accepts the current scene variables and the advertising content encoding as input. After that, the system calculates the information entropy after delivery: ; Where, Indicates simulated advertising After that, users react to the advertisement The probability of selection; Represents the information entropy in the simulation state; is the total number of candidate ads. Then the information entropy change Measuring Advertising The impact of delivery: ; Where, Indicates candidate ads The information entropy change value caused by the placement of the advertisement selection structure; if , indicating that the advertisement will guide the user to focus more on the advertisement interest, which is beneficial to interest focusing; The information entropy selected for the advertisement.

[0059] In some embodiments, to avoid the impact of invalid advertisements on user experience, the system introduces an information entropy change threshold , and only when a certain advertisement meets , it can be included in the candidate set. Finally, from the advertisements that meet the conditions, the one with the largest entropy drop amplitude is selected, that is: ; In the formula, is the finally determined target advertisement for delivery; is any th advertisement in the candidate advertisement set.

[0060] As an option, the above process can be completed in an edge computing node, and the information entropy calculation and the prediction distribution generation process can be accelerated by tensor batch parallel, significantly reducing the response time delay.

[0061] Specifically, in the new energy vehicle charging scene, if the user's historical behavior shows that he prefers entertainment type advertisements on weekends, the system can analyze the information entropy drop curve to preferentially select the advertisement item that makes the advertisement response focus on this type of content, rather than blindly delivering the current highest scoring advertisement, avoiding generalized recommendation.

[0062] For step S5, in this embodiment, after the display of the target advertisement is completed, the user behavior will respond to the advertisement, such as clicking, jumping or browsing and staying, etc. These behaviors constitute the first-hand feedback data of the effect of the advertisement strategy. The actual response behavior of the user is used to correct the user factor vector generated in the tensor decomposition model in reverse, so as to realize the dynamic update and fine adjustment of the user portrait.

[0063] Generally, the system will record the user's interactive behavior data after a round of advertisement delivery. The behavior data includes but is not limited to the following information: Whether the user clicks the advertisement; Whether to jump to an external link after clicking; Page stay time; Whether the user closes the advertisement or returns to the previous page; Whether other path interactions (such as secondary clicks, collections, evaluations, etc.) are generated during browsing.

[0064] In one possible implementation, the system combines the above behavior data to generate a response vector , where represents the response feature dimension. Each dimension can correspond to a specific behavioral quantification indicator, such as click = 1, no click = 0, dwell time normalized to [0, 1], etc.

[0065] The advertisement itself also has a structured representation, and the embedding vector of the advertisement that has been launched is denoted as , which is consistent with the aforementioned user factor vector .

[0066] In order to learn the deviation between the real behavior of the user and the predicted behavior of the model, the system defines the following loss function: ; In the formula, is the response vector of the user to the launched advertisement; is the embedding vector of the advertisement; is the user factor vector; is the loss function, represents a matching function for predicting the user response; in one implementation, the matching function is defined as: ; In the formula, is a Sigmoid function, used to normalize the matching result to [0, 1]; represents the similarity score between the user and the advertisement vector; represents a matching function for predicting the user response; is the inner product operation between the user and the advertisement vector.

[0067] After completing the construction of the loss function, the system calculates the gradient of the loss function with respect to the user factor vector , to obtain the update gradient for correcting the vector , which is as follows: ; In the formula, is the gradient of the loss function with respect to the user factor vector ; the partial derivative of the loss function with respect to ; is the predicted advertisement response; is the derivative of the activation function ; this gradient is used to update , which can be executed in one step correction as follows: ; In the formula, is the learning rate, which controls the update amplitude.

[0068] In some embodiments, when the response behavior is complex, a multi-dimensional label model can be used to extend the loss function to a weighted multi-task loss. For example, a joint loss is constructed considering click-through rate (CTR), conversion rate (CVR), and dwell time (DwellTime) at the same time: ; In the formula, is the total loss function value; is the loss function item of click-through rate prediction; is the conversion rate prediction loss item; is the dwell time loss function item.

[0069] In the above loss function combination, is an empirical weight coefficient, which can be set by cross-validation.

[0070] In the updated vector is synchronized to the user portrait module, which is used as the basis input for the next round of advertisement scoring and strategy generation, ensuring that the expression of user preferences can be continuously corrected with actual interaction behavior, and having dynamic response capability.

[0071] For step S6, in this embodiment, with the continuous delivery of advertisements and the continuous evolution of user behavior, relying only on the initial tensor modeling results for strategy matching may cause problems such as portrait staticization and model aging. A cloud-edge collaborative dynamic feedback mechanism is introduced to drive model retraining through user behavior changes, realizing periodic adaptive update of advertisement strategies.

[0072] First, the updated user behavior embedding vector is analyzed for changes locally (i.e., at the edge device), and a simplified representation is generated through structural compression as feedback information required for cloud training.

[0073] Specifically, let the user factor vector before updating be and the updated one be Then the system can calculate the vector difference as: ; In the formula, represents the change of the user embedding vector after one advertisement response behavior or model update.

[0074] As an option, to compress the transmission bandwidth and improve the semantic concentration of the abstract, the system can further extract the vector change direction and intensity as a feature abstract: Vector change amplitude: ; Change direction (unit vector normalization can be used); Change time and behavior type label.

[0075] In a possible implementation, the summary constitutes the following structure: ; In the formula, is a behavior vector change summary of the user ; is a timestamp; is a scene identifier; is a response type identifier.

[0076] The summary information will be uploaded to the cloud host server as feedback data for the construction of a training sample set.

[0077] After the cloud system receives the user change summaries from multiple edge nodes, it relies on the same to construct a fine-tuning sample set of the behavior tensor model . These samples can be regarded as implicit preference change signals for reconstructing the latent factor parameter set .

[0078] As an implementation, the tensor modeling adopts a CP decomposition structure, and the objective function is: ; In the formula, is a tensor reconstruction loss function; is an index set of non-zero observation data; is an original observation value in the third-order behavior tensor; are respectively the ranks of the user, time, and location dimensions components; is the rank of the tensor decomposition, indicating the number of hidden space dimensions; denotes the Frobenius norm, used for regularization constraints; is a regularization coefficient.

[0079] In the fine-tuning process, the system not only uses the tensor reconstruction error as the objective function, but also can optimize performance indicators such as advertisement click-through rate (CTR) and conversion rate (CVR) in parallel, forming a multi-objective optimization structure, so as to balance the prediction accuracy and behavior matching efficiency in model updating.

[0080] After training is completed, the system performs structural encoding on the latest round of latent factor parameters and distributes them to the edge nodes in the form of modules. Specifically, the distribution content can include: vectors corresponding to the active user set; time segment embedding matrix and geographic position embedding matrix; strategy adjustment parameters, such as recommendation factor weight, entropy constraint threshold, etc.

[0081] After the edge node receives the parameters, vector replacement or parameter interpolation operations in the strategy module can be automatically completed to complete the distributed collaborative evolution of the local model.

[0082] In some embodiments, to improve the robustness of the system, a version control mechanism can also be introduced, and each round of parameter distribution includes a model number to prevent repeated loading or concurrent conflicts of old parameters.

[0083] In some embodiments, to improve the robustness of the system, a version control mechanism can also be introduced, and each round of parameter distribution includes a model number model_ver to prevent repeated loading or concurrent conflicts of old parameters.

[0084] The charging pile-based advertisement delivery system described below can correspond to the charging pile-based advertisement delivery method described above.

[0085] Please refer to the attached Figure 2 The application also provides a charging pile-based advertisement delivery system, comprising: A data acquisition module is configured to acquire various types of data, including user behavior data, device state data, scene environment data, and advertisement response data. A tensor modeling module is configured to construct a third-order behavior tensor based on the data provided by the data acquisition module. A strategy optimization module is configured to calculate the probability distribution of user acceptance of advertisements based on the generated user behavior embedding vector, and to use information entropy to evaluate the potential effect of advertisement acceptance to generate an optimal delivery strategy. A strategy screening module is configured to screen the optimal candidate advertisement based on the optimal delivery strategy and the change amount of information entropy. An edge delivery module is configured to perform actual pushing of advertisements at the charging pile end and to evaluate the effect of the advertisements through collected feedback information. A strategy updating module is configured to update the parameters of the delivery strategy after receiving the feedback information from the edge delivery module, and to synchronize the updated strategy to the devices at the charging pile end.

[0086] The system of the embodiment can be used to perform the above-mentioned method embodiments, and has similar principles and technical effects, which will not be described here.

[0087] The embodiments of the specific implementation are preferred embodiments of the application, and are not limited to the protection scope of the application, wherein the same parts are denoted by the same reference numerals. Therefore, any equivalent changes made according to the structure, shape, principle of the application should be covered within the protection scope of the application.

Claims

1. A charging pile-based advertising method, characterized in that: The following steps are involved: S1. Collect interaction data between users and charging piles, user behavior data, location information, time information and environmental status data to build a data set of behavior characteristics; S2. Based on the behavioral feature data set, a third-order behavior tensor is constructed with user ID, time segment, and geographic location as dimensions. The third-order behavior tensor is then decomposed to generate the corresponding user behavior embedding vector. S3. Match the user behavior embedding vector with the candidate advertisement set, establish a probability distribution model of the user's response to each candidate advertisement, and calculate the information entropy value of the advertisement selection based on the probability distribution model of each candidate advertisement; S4. Identify and select the target advertisement that causes the largest entropy decrease based on the information entropy value, and select it as the current advertisement; S5. Display the target advertisement on the charging pile terminal, collect user response behavior data to the target advertisement, and update user behavior characteristics based on the response behavior; S6. Feedback the updated user behavior features to the cloud system for retraining the behavior tensor model and dynamically adjusting the advertising delivery strategy.

2. The method for placing advertisements based on charging piles according to claim 1, characterized in that: In step S1, constructing a data set of behavioral features includes the following steps: Based on the collected interaction data between users and charging stations, a user-time interaction mapping table is constructed, and by matching location information with weather data, a behavioral context with spatiotemporal scene semantics is generated; The behavior context drives the construction of a multidimensional data structure, where each behavior record is mapped into a structured vector containing user ID, charging duration, time tag, weather status, and scene number; The structured vector serves as the data basis in the subsequent tensor modeling step and is input into the index field of the tensor dimension generator to define the behavior tensor.

3. The method for placing advertisements based on charging piles according to claim 1, characterized in that: In step S2, generating the corresponding user behavior embedding vector includes the following steps: Based on the generated structured vector, the user number, time segment number and geographical area number are used as the three dimensions of the third-order behavior tensor. In the tensor construction process, the existence or non-existence of behavioral events is used as the criterion for setting the tensor value, forming a third-order sparse behavior tensor; The CP decomposition method that minimizes the reconstruction error is used to decompose the tensor and generate the user embedding vector.

4. The method for placing advertisements based on charging piles according to claim 3, characterized in that: The tensor is decomposed by the CP decomposition method that minimizes the reconstruction error. The minimization of the reconstruction error adopts the following formula: ； Where, Represents interactive behavior data of users, time and location, Represent the rank-one components in the user, time, and location dimensions respectively, represents the dimension of the latent space, Represents the vector outer product.

5. The method for placing advertisements based on charging piles according to claim 1, characterized in that: In step S3, the information entropy value of advertisement selection is calculated based on the probability distribution model of each candidate advertisement, including the following steps: The user embedding vector is input into the matching scoring network and the inner product operation is performed with each ad feature vector to generate a matching score; All ad scores are normalized by Softmax to obtain the probability of the user selecting each ad. The Softmax normalization calculation formula is: ； Where, represents the probability that the user selects the ad. Represents the exponential sum of all advertisement scores. No. The score of an ad, is the base of natural logarithms; Construct information entropy function based on probability distribution , the information entropy is used to quantify the distribution of current advertisements; Where, Information entropy selected for the ad, represents the probability that the user selects the ad. For the The logarithm of the probability of selecting an ad, is the total number of ads.

6. The method for placing advertisements based on charging piles according to claim 1, characterized in that: In step S4, identifying and screening the target advertisement that causes the largest entropy decrease based on the information entropy value includes the following steps: Call the established probability distribution model to simulate the changes in user behavior after each candidate ad is delivered, and generate a predicted probability distribution; For each candidate advertisement, calculate the change in information entropy as the change in information structure after the advertisement is released; Based on the information entropy change results of all candidate ads, the ads with the largest information entropy decrease are screened and the constraints that their information entropy decrease must meet are set as the final ads to be released.

7. The method for advertising based on charging piles according to claim 1, characterized in that: In step S5, updating the user behavior characteristics based on the response behavior includes the following steps: Combine the display records of the delivered ads with the user clicks, browsing time, and interaction paths to generate a response vector; Calculate the update gradient based on the response vector and the delivered ad vector, and use it to reversely correct the user factor vector in the tensor decomposition model; Through the modified user factor vector, the user's behavioral preference changes in the third-order tensor are dynamically mapped.

8. The method for placing advertisements based on charging piles according to claim 7, characterized in that: The calculation of the update gradient based on the response vector and the delivery advertisement vector includes the following steps: Set the user's response vector to the delivered ad to , the embedding vector of the advertisement is , the current user factor vector is ; The loss function constructed to calculate the user behavior preference error is: ； Where, is the loss function, represents the matching function for predicting user responses; By calculating the loss function about Derivative, get the updated gradient of the user factor vector: ； Where, is the loss function for the user factor vector The gradient, Loss function pair The partial derivative of For the predicted ad response, is the activation function The derivative of .

9. The method for placing advertisements based on charging piles according to claim 1, characterized in that: In step S6, the retraining of the behavior tensor model and the dynamic adjustment of the advertising delivery strategy include the following steps: Summarize the updated user embedding vector changes, extract the feature data of the vector change direction and magnitude, and upload it to the cloud; The cloud system uses the change summaries uploaded by multiple edge nodes to construct a sample set for tensor model fine-tuning and retrains the latent factor parameter set based on the sample set; After training is completed, the updated parameters are sent to the edge devices to complete the distributed collaborative evolution of the policy model.

10. An advertising delivery system based on a charging pile, applied to an advertising delivery method based on a charging pile according to any one of claims 1 to 9, characterized in that: include: Data collection module, used to collect various types of data, including user behavior data, device status data, scene environment data and advertising response data; A tensor modeling module is used to construct a third-order behavior tensor based on the data provided by the data acquisition module; The strategy optimization module is used to calculate the probability distribution of users accepting ads based on the generated user behavior embedding vector, and use information entropy to evaluate the potential effect of ad acceptance and generate the optimal delivery strategy; The strategy screening module is used to screen out the best candidate ads based on the optimal delivery strategy and the change in information entropy; The edge delivery module is used to push ads to charging stations and evaluate the effectiveness of ads based on collected feedback. The strategy update module is used to update the parameters of the delivery strategy after receiving feedback information from the edge delivery module, and synchronize the updated strategy to the device at the charging pile end.

Images