A method and system for advertising delivery based on charging piles

Abstract

This application relates to the field of new energy charging and advertising technology, and discloses an advertising delivery method and system based on charging piles. The method includes: collecting user interaction, behavior, location, time, and environmental data to construct a behavioral feature dataset; constructing a third-order behavioral tensor based on the behavioral features and decomposing it to generate user behavior embedding vectors; calculating the information entropy value of the selected advertisement; selecting the target advertisement with the largest decrease in entropy for delivery; displaying the target advertisement, collecting user responses, and updating behavioral features; and retraining the behavioral tensor model to adjust the advertising strategy. The system includes: a data acquisition module, a tensor modeling module, a strategy optimization module, a strategy selection module, an edge delivery module, and a strategy update module. This invention achieves high-dimensional semantic modeling of user behavior by constructing a behavioral tensor with user identifier, time segment, and geographical location as three-dimensional indices, and introducing a tensor decomposition method to extract behavioral embedding vectors.

Description

Technical Field

[0001] This application relates to the field of new energy charging and advertising technology, specifically to an advertising method and system based on charging piles. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the density of urban public charging stations is constantly increasing, gradually evolving from "infrastructure" to "interactive platform." During the charging process, users' attention is relatively focused and their dwell time is controllable, objectively creating a natural scenario for advertising display. These scenarios possess strong physical location attributes, clear time-period characteristics, and user behavior characteristics highly correlated with vehicle usage habits, thus they are widely considered to have high potential for information reach.

[0003] Existing advertising technologies largely rely on existing mobile advertising frameworks for deployment. The basic idea is to use charging devices as advertising terminals and statically match users' basic identity information (such as account, vehicle type, region, etc.) with preset advertising tags. The advertising selection logic often depends on preset rules, targeting strategies, or simple classification algorithms for delivery. The content is mostly based on a uniform advertising package set by region, time, or the service provider's backend, lacking in-depth analysis of the dynamic behavioral characteristics of individual users.

[0004] However, existing advertising technologies rely too heavily on static tags for user profiling, making it difficult to integrate multi-dimensional data such as location information, charging behavior, and environmental variables, resulting in simplistic and rigid user profiles. Secondly, advertising strategies lack the ability to dynamically capture evolving user behavior, making it difficult to adjust content distribution paths in a timely manner based on user feedback. Thirdly, traditional recommendation model training mechanisms generally employ a centralized, periodic update model, resulting in low update frequency and slow response. Therefore, this invention provides an advertising delivery method and system based on charging stations to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide an advertising delivery method and system based on charging piles, which solves the problems of single user profile dimensions, delayed behavioral response, and slow strategy updates in existing advertising delivery technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an advertising placement method based on charging piles, comprising the following steps:

[0007] S1. Collect user interaction data with charging piles, user behavior data, location information, time information, and environmental status data to construct a dataset of behavioral characteristics;

[0008] S2. Based on the dataset of behavioral features, construct a third-order behavioral tensor with user identifier, time segment and geographical location as dimensions, and decompose the third-order behavioral tensor to generate the corresponding user behavior embedding vector;

[0009] S3. Match the user behavior embedding vector with the candidate ad set, establish a probability distribution model of the user for each candidate ad, and calculate the information entropy value of ad selection based on the probability distribution model of each candidate ad.

[0010] S4. Identify and filter the target advertisement that causes the largest decrease in entropy based on the information entropy value, and use it as the current advertisement to be delivered;

[0011] 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;

[0012] S6. Feed the updated user behavior features back to the cloud system for retraining the behavior tensor model and dynamically adjusting the advertising strategy.

[0013] Preferably, in step S1, constructing the dataset of behavioral features includes the following steps:

[0014] Based on the collected user-charging pile interaction data, 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.

[0015] The behavior context drives the construction of a multidimensional data structure, where each behavior record is mapped to a structured vector containing user identifier, charging duration, time tag, weather status, and scene number;

[0016] The structured vector serves as the data foundation for subsequent tensor modeling steps and is input into the tensor dimension generator to define the index field of the behavioral tensor.

[0017] Preferably, in step S2, generating the corresponding user behavior embedding vector includes the following steps:

[0018] Based on the generated structured vector, the user ID, time segment ID, and geographic region ID are used as the three dimensions of the third-order behavior tensor.

[0019] During the tensor construction process, the existence of behavioral events is used as the criterion for setting tensor values, thus forming a third-order sparse behavioral tensor.

[0020] The tensor is decomposed using the CP decomposition method that minimizes the reconstruction error to generate user embedding vectors.

[0021] Preferably, the tensor is decomposed using the CP decomposition method that minimizes the reconstruction error, wherein the minimization of the reconstruction error is achieved using the following formula:

[0022] ;

[0023] In the formula, Data representing user, time, and location interaction behaviors. These represent the rank-one components in the user, time, and location dimensions, respectively. The dimension of the latent space is represented. This represents the outer product of vectors.

[0024] Preferably, in step S3, calculating the information entropy value of ad selection based on the probability distribution model of each candidate ad includes the following steps:

[0025] 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.

[0026] The probability of a user selecting each advertisement is obtained by normalizing all advertisement scores using Softmax. The Softmax normalization formula is as follows:

[0027] ;

[0028] In the formula, This indicates the probability that a user will select the advertisement. This represents the sum of the scores for all advertisements. No. The score of each advertisement, is the base of the natural logarithm;

[0029] Construct an information entropy function based on probability distribution. The information entropy is used to quantify the distribution of the current advertisements;

[0030] In the formula, The information entropy chosen for advertising This indicates the probability that a user will select the advertisement. For the first The logarithm of the probability of choosing an advertisement. This represents the total number of advertisements.

[0031] Preferably, in step S4, identifying and filtering the target advertisement that causes the largest decrease in entropy based on the information entropy value includes the following steps:

[0032] The established probability distribution model is invoked to simulate the changes in user behavior after each candidate ad is delivered, and the predicted probability distribution is generated.

[0033] For each candidate ad, calculate the change in information entropy, which is used as the change in information structure after the ad is delivered.

[0034] Based on the information entropy changes of all candidate ads, the ad with the largest decrease in information entropy is selected and a constraint is set for its information entropy decrease to be met, which is then used as the final ad for delivery.

[0035] Preferably, in step S5, updating user behavior features based on response behavior includes the following steps:

[0036] The display records of the ads that have been displayed are combined with user clicks, browsing time and interaction paths to generate response vectors;

[0037] The updated gradient is calculated based on the response vector and the ad delivery vector, and is used to correct the user factor vector in the tensor decomposition model in reverse.

[0038] The modified user factor vector dynamically maps changes in user behavioral preferences in the third-order tensor.

[0039] Preferably, the calculation of the update gradient based on the response vector and the delivered ad vector includes the following steps:

[0040] Let the user's response vector to the displayed advertisement be set as... The embedding vector of the advertisement is The current user factor vector is ;

[0041] The loss function for calculating user behavior preference error is constructed as follows:

[0042] ;

[0043] In the formula, For loss function, This represents a matching function that predicts user responses.

[0044] By calculating the loss function with respect to Take the derivative to obtain the update gradient of the user factor vector:

[0045] ;

[0046] In the formula, For the loss function on the user factor vector gradient, loss function pairs The partial derivatives, For predicted advertising response, Activation function The derivative of .

[0047] Preferably, in step S6, the retraining of the behavioral tensor model and dynamic adjustment of the advertising delivery strategy includes the following steps:

[0048] The updated user embedding vector changes are summarized, and feature data of the vector change direction and magnitude are extracted and uploaded to the cloud.

[0049] The cloud system uses change summaries uploaded from multiple edge nodes to construct a sample set for fine-tuning the tensor model, and then retrains the latent factor parameter set based on this sample set.

[0050] After training is complete, the updated parameters are sent to edge devices to complete the distributed collaborative evolution of the policy model.

[0051] An advertising delivery system based on charging piles is also provided, including:

[0052] The data acquisition module is used to collect various types of data, including user behavior data, device status data, scene environment data, and advertising response data.

[0053] The tensor modeling module is used to construct a third-order behavioral tensor based on the data provided by the data acquisition module.

[0054] The strategy optimization module is used to calculate the probability distribution of users accepting ads based on the generated user behavior embedding vectors, and to use information entropy to evaluate the potential effect of ad acceptance and generate the best delivery strategy.

[0055] The strategy filtering module is used to filter out the best candidate ads based on the change in information entropy, according to the best delivery strategy.

[0056] The edge delivery module is used to actually push advertisements at the charging pile and evaluate the advertising effect through the collected feedback information;

[0057] The strategy update module is used to receive feedback information from the edge delivery module, update the parameters of the delivery strategy, and synchronize the updated strategy to the devices at the charging pile.

[0058] In summary, this application includes at least one of the following beneficial technical effects:

[0059] 1. This invention constructs a behavior tensor with user identifier, time segment, and geographic location as three-dimensional indices, and introduces tensor decomposition to extract behavior embedding vectors, thereby achieving high-dimensional semantic modeling of user behavior. Compared to traditional advertising systems that rely solely on static indicators such as click-through rate and tagged user features, making it difficult to capture deep preference associations, this method more accurately depicts the evolution path of user interests, effectively compensating for the technical shortcomings of coarse-grained user profiles and singular behavior modeling.

[0060] 2. This invention uses information entropy reduction as the criterion for selecting target advertisements. Instead of relying solely on matching scores for ranking during the advertisement selection process, it introduces a structural indicator of user selection uncertainty. This approach breaks free from the limitations of linear recommendation frameworks, enabling dynamic identification of advertisement content that truly possesses "information focus" capabilities. It solves the problems of low click-through rates and low conversion rates caused by previous recommendation systems that easily pushed redundant and distracting ad segments.

[0061] 3. This invention introduces a mechanism for updating user factor vectors after ad response and performs gradient backpropagation by constructing a loss function and the response vector. The user profile possesses the ability to continuously learn and improve with ad interaction. Unlike existing technologies that rely on periodic full retraining, this scheme achieves rapid adaptation at the edge level, solving the efficiency bottleneck of delayed response in behavioral modeling.

[0062] 4. This invention combines user embedding vector variation summaries uploaded from multiple edge devices, enabling the system to fine-tune and train a tensor model in the cloud, and achieve synchronous collaboration of the model across edge nodes through parameter distribution. This design abandons the high-cost model of traditional centralized large-scale global updates, achieving distributed and lightweight optimization. Especially in application environments with changing scenarios and real-time user behavior, it avoids recommendation failures caused by model lag, significantly improving the system's adaptability and robustness in actual operation. Attached Figure Description

[0063] Figure 1 This is a flowchart of the method steps in this application;

[0064] Figure 2 This is the system architecture diagram of this application. Detailed Implementation

[0065] The following is in conjunction with the appendix Figure 1 -Appendix Figure 2 This application will be described in further detail below.

[0066] Please see the appendix Figure 1 This invention provides a method for a multimodal retrieval engine for security personnel and vehicles, comprising the following steps:

[0067] S1. Collect user interaction data with charging piles, user behavior data, location information, time information, and environmental status data to construct a dataset of behavioral characteristics;

[0068] S2. Based on the dataset of behavioral features, construct a third-order behavioral tensor with user identifier, time segment and geographical location as dimensions, and decompose the third-order behavioral tensor to generate the corresponding user behavior embedding vector;

[0069] S3. Match the user behavior embedding vector with the candidate ad set, establish a probability distribution model of the user for each candidate ad, and calculate the information entropy value of ad selection based on the probability distribution model of each candidate ad.

[0070] S4. Identify and filter the target advertisement that causes the largest decrease in entropy based on the information entropy value, and use it as the current advertisement to be delivered;

[0071] 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;

[0072] S6. Feed the updated user behavior features back to the cloud system for retraining the behavior tensor model and dynamically adjusting the advertising strategy.

[0073] For step S1, in this embodiment, to overcome the problems of coarse user profiles and single tag dimensions in existing advertising delivery systems in charging pile scenarios, a behavior data-driven profile modeling mechanism is introduced at the charging pile end. This involves collecting direct interaction information between users and devices, focusing on auxiliary factors such as environmental status, time period, and spatial distribution. This forms a behavioral feature data set with spatiotemporal semantics, behavioral context, and user identifier fusion capabilities.

[0074] When a charging action occurs, the system prioritizes establishing a user-time interaction mapping relationship and constructing an original behavior event table. For example, if a user U charges at a commercial charging station B at 18:45 on March 15, 2025, their charging start and end times, device ID, power usage, etc., will be collected in real time.

[0075] As one possible implementation, this behavioral event is associated with a location information database to obtain the encoded information L of the geographical area to which the charging pile belongs. Simultaneously, the system calls an external interface to obtain the weather data corresponding to that time point, including temperature, weather type (sunny / rainy / snowy, etc.), and humidity level, generating behavioral context fields.

[0076] Specifically, this behavioral context can be encoded as the following structured entries:

[0077] User identifier (e.g., U number);

[0078] Charging start time, end time, and duration;

[0079] Geographic location code (e.g., GPS / business district code);

[0080] Time tags (e.g., weekday / weekend, morning rush hour / evening rush hour);

[0081] Weather status codes (e.g., 01 indicates sunny, 03 indicates light rain);

[0082] Scene number (optional, selected from the defined scene dictionary).

[0083] In one possible approach, the above structure is further processed into the indexed form required for a multidimensional tensor input format. Constructing a three-dimensional tensor. The meanings of each dimension are as follows:

[0084] This represents a set of user IDs, with a range of IDs. , ;

[0085] Represents a set of time segments, divided by hours or natural blocks;

[0086] This represents a set of geographic location codes, defined by region or business district.

[0087] Each structured behavior vector can be mapped to a non-zero element in a tensor, whose value can be set as a behavior intensity index (such as charging time, interaction frequency, etc.) or a Boolean value. This indicates that an event has occurred. This tensor will be used as the underlying data structure for latent factor modeling in subsequent steps.

[0088] As an alternative, to enhance the timeliness and update capability of data, the system can be configured with a sliding window mechanism. For example, the user behavior mapping set could be refreshed every 24 hours, retaining only the most recent data. This is an effective action to avoid map inflation and timeliness distortion.

[0089] In practical applications, such as when users exhibit different charging preferences under different weather conditions (for example, they tend to charge in the underground parking lot of a shopping mall on rainy days), the behavioral context can reflect this potential behavioral variability and provide an important basis for subsequent advertising scenario matching.

[0090] In some implementations, the system can additionally introduce interfaces to connect with operator data to collect implicit features such as user communication activity, movement trajectory, and APP usage frequency, further enriching the user profile input space. Subject to data authorization, POS payment records and consumption preference types can also be introduced as auxiliary tags to participate in multi-dimensional modeling.

[0091] The constructed dataset not only serves as the basis for subsequent tensor structure generation but also as input for training a deep behavior prediction model. During model training, this dataset is fed into the user behavior preference model in batches, and the model optimizes itself by minimizing the difference between predicted and actual behaviors.

[0092] For example, the system can introduce the following benchmark loss function in the early stages of construction:

[0093] ;

[0094] In the formula, This represents the total number of data points. For the first behavioral data, The model predicts the values, and the loss function measures the error in reconstructing the behavioral features. The smaller the value, the better the data structure restoration effect.

[0095] In general, the system will perform user segmentation and labeling at this stage, dividing users into behavioral clusters such as "highly active", "periodic", and "weekend consumer" to guide the subsequent ad matching model to adopt different strategies.

[0096] For step S2, in this embodiment, after constructing the user behavior feature data set, it is necessary to further map the 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 extracts the latent factor features implicit in the spatiotemporal behavior dimension through tensor modeling and decomposition methods, and finally generates an embedded representation of user behavior as the input basis for the subsequent advertising strategy model.

[0097] In this embodiment, the main components include: the construction of a third-order behavioral tensor and the generation of user embedding vectors based on CP decomposition.

[0098] Specifically, the system first encodes the structured identifiers in user behavior events:

[0099] Encode the user ID as a user dimension index ;

[0100] Encode time segment information into a time dimension index It can be divided by hour, intraday cycle or behavior block;

[0101] Map geographic region codes to spatial location dimension indices It can be set according to GPS grid, administrative region or business district number.

[0102] In some embodiments, whether a behavioral event has occurred can be labeled as a Boolean variable, and the following tensor value setting criteria are defined:

[0103] If in the index A user action event occurred at the corresponding location (e.g., the user was in a time slice). Within the region (the charging behavior), then the third-order tensor ,otherwise, .

[0104] In general, this tensor is highly sparsity, with non-zero values ​​assigned only at points of actual user behavior.

[0105] In the above tensor After construction, the system is modeled using the CP decomposition method, which minimizes the reconstruction error. This method represents the tensor as the sum of the outer products of multiple rank tensors, aiming to approximate the original behavioral tensor as the product of three sets of low-rank embedding matrices. Its optimization objective function is as follows:

[0106] ;

[0107] In the formula, Data representing user, time, and location-based interaction behaviors; These represent the rank-one components in the user, time, and location dimensions, respectively. Represents the outer product of vectors; parameters The rank of the decomposition is usually a hyperparameter, and its value is determined by actual performance tuning. This represents the Frobenius norm, used to measure the error in tensor approximation reconstruction.

[0108] In one possible implementation, the system randomly initializes the latent factor matrix and iteratively solves the minimization problem using methods such as Alternating Least Squares (ALS) or the Adam optimizer.

[0109] Ultimately, the tensor decomposition will generate a set of embedding vectors, where each user corresponds to a set of vectors representing their behavioral preference features in the latent semantic space. These embedding vectors will serve as one of the inputs to the subsequent ad matching and scoring model, playing a role in characterizing users' long-term preferences and behavioral features.

[0110] In some extended implementations, to enhance expressive power, the system can also embed the tensor modeling process into the neural network structure to build a deep tensor decomposition model (such as TensorFusionNetwork). The embedded parameters and policy network are simultaneously optimized through end-to-end training, thereby enhancing the model's ability to fit complex behavioral patterns.

[0111] It is worth noting that user embedding vectors not only contain statistical features of their historical behavior, but also implicitly contain joint preferences for the temporal and spatial dimensions. This enables subsequent policy models to more accurately identify users' response tendencies in specific scenarios when dealing with dynamic scene matching.

[0112] For step S3, in this embodiment, after generating the user behavior embedding vector, in order to achieve accurate filtering and ranking of personalized advertisements, the system needs to establish a prediction model of the user's response probability to the candidate advertisement set based on the embedding vector. This model is not only used to estimate the user's tendency to click on or follow specific advertisements, but also provides numerical basis for information entropy evaluation and target advertisement filtering strategies.

[0113] An advertising probability distribution prediction function is constructed, and the uncertainty of advertising selection is quantified based on this distribution. This process comprehensively considers the expressive power of advertising feature vectors and the semantic matching degree between user behavior embedding vectors.

[0114] Typically, the system maintains a set of feature vectors containing multiple ad candidates, denoted as . , where each vector Indicates the first One candidate ad in Feature representation in a 3D semantic space.

[0115] User behavior embedding vector This originates from the aforementioned tensor decomposition process. The system will... The input is fed into the scoring and matching module, where it is matched against all candidate ad vectors. Alternatively, the matching function can be a dot product, i.e.:

[0116] ;

[0117] In the formula, For users to advertise The matching score is the user's score. A higher score indicates a stronger preference for the ad. After calculating the matching score, the system sets the scores together. The input is fed into the Softmax function, which normalizes it into a probability distribution to obtain the probability of a user selecting each advertisement. The expression for the Softmax function is as follows:

[0118] ;

[0119] In the formula, This indicates the probability that a user will select the advertisement. This represents the sum of the scores for all advertisements. No. The score of each advertisement; is the base of the natural logarithm.

[0120] In one possible implementation, the system will select all ad candidates. Composition probability distribution This distribution not only reflects users' tendency to respond to advertisements in the current scenario, but also indirectly represents the distribution density of behavioral preferences.

[0121] To measure the determinism of the ad response distribution, the system further introduces an information entropy metric to characterize the structural complexity of ad selection. A higher entropy value indicates a more dispersed distribution of user interest across ads; a lower entropy value indicates more focused user interest and a clearer ad preference. The information entropy function is defined as follows:

[0122] ;

[0123] in, The entropy value represents the probability distribution of the current ad selection. It is a logarithmic function, usually with the natural logarithm as its base; User's The probability of selecting each advertisement is derived from the Softmax normalization result; This represents the total number of advertisements.

[0124] As an alternative, to improve the real-time performance and scalability of the system, probability calculation and entropy evaluation can be performed in parallel, and some model parameters can be preloaded in edge computing nodes to achieve lightweight deployment.

[0125] In some embodiments, the system may also consider the impact of ad display order by incorporating ad context factors. right Perform dynamic adjustments and construct a context-aware matching score function:

[0126] ;

[0127] In the formula, For users to advertise Match score; To calculate the inner product between the user vector and the ad correction vector; For users The behavioral feature embedding vector.

[0128] The above processing can further improve the model's adaptability to different scenarios, and is particularly suitable for charging scenarios with complex ad placement layouts or sensitive user behavior timing.

[0129] For step S4, in this embodiment, given that the user-ad matching probability distribution has been established, relying solely on the ad with the highest probability for direct delivery may lead to repetitive pushes or dilution of interest. To further improve the matching stability between ad content and user status, an ad filtering strategy driven by information entropy changes is introduced.

[0130] By quantifying the structural changes in users' probabilities of choosing ads before and after ad delivery, the system can dynamically identify which candidate ads can effectively reduce the uncertainty of users' decisions. Finally, using the magnitude of entropy reduction as the screening criterion, the system determines the target ads with the most structurally beneficial delivery results.

[0131] In this embodiment, the system is based on the advertising probability distribution obtained in the previous stage. First, we construct its information entropy function to represent the current uncertainty state of the advertising response distribution:

[0132] ;

[0133] In the formula, Information entropy selected for advertising; This indicates the probability that a user will select the advertisement. For the first The logarithm of the probability of selecting an advertisement; This represents the total number of advertisements.

[0134] Subsequently, the system processed each candidate advertisement. Conduct a campaign simulation. Specifically, simulate the probability distribution changes in user behavior after the ad is launched, denoted as . This distribution can be achieved by introducing a delivery response offset factor. For example, it can be used to predict potential shifts in preferences after delivery based on users' past response models.

[0135] In one possible implementation, the system introduces a prediction module to predict the performance of a given advertisement. Estimate click-through rate, dwell time, or conversion rate to infer... The estimation form is as follows. This prediction module can be constructed from a lightweight neural network, accepting current scene variables and ad content encoding as input. The prediction distribution is then obtained. Then, the system calculates the post-deployment information entropy based on this:

[0136] ;

[0137] In the formula, Indicates simulated advertising. Afterwards, users reacted to the advertisements. The probability of choosing; This represents the information entropy under simulated conditions. This represents the total number of candidate ads. Then, the change in information entropy... Measuring advertising Impact of deployment:

[0138] ;

[0139] In the formula, Indicates candidate ads The change in information entropy caused by the placement of ads to the ad selection structure; if This indicates that placing this ad will guide users to focus their interest on the ad more effectively, which is beneficial for focusing their interest. Information entropy selected for advertising.

[0140] In some embodiments, to prevent invalid ads from affecting user experience, the system introduces an information entropy change threshold. and only if an advertisement satisfies Only when the condition is met can it be included in the candidate set. Finally, from the advertisements that meet the conditions, the one with the largest decrease in entropy is selected, i.e.:

[0141] ;

[0142] In the formula, For the finalized target ads; For any 1 in the candidate ad set An advertisement.

[0143] Alternatively, the above process can be completed in edge computing nodes, and the information entropy calculation and prediction distribution generation process can be accelerated by tensor batch parallelism, significantly reducing response latency.

[0144] Specifically, in the scenario of charging new energy vehicles, if a user's historical behavior shows that they prefer entertainment-related ads on weekends, the system can analyze the information entropy decline curve and prioritize ads that concentrate the ad response on that type of content, rather than simply displaying the ads with the highest current scores, thus avoiding generalized recommendations.

[0145] In step S5, in this embodiment, after the target advertisement is displayed, user behavior will respond to the advertisement, such as clicking, jumping, or browsing. These behaviors constitute first-hand feedback data on the effectiveness of the advertising strategy. The actual user response behavior is used to reverse-correct the user factor vector already generated in the tensor decomposition model, thereby achieving dynamic updating and fine-tuning of the user profile.

[0146] Generally, the system records user interaction data after a round of ad delivery. This interaction data includes, but is not limited to, the following information:

[0147] Did the user click on the ad?

[0148] Does clicking this link redirect to an external link?

[0149] Page dwell time;

[0150] Does the user want to close the ad or return to the previous page?

[0151] Does the browsing process involve other interactive paths (such as secondary clicks, favorites, ratings, etc.)?

[0152] In one possible implementation, the system combines the aforementioned behavioral data to generate a response vector. ,in This represents the response feature dimension. Each dimension can correspond to a specific behavioral quantification indicator, such as click = 1, no click = 0, and dwell time normalized to [0, 1], etc.

[0153] The advertisement itself also has a structured representation; let the embedding vector of the delivered advertisement be... With the aforementioned user factor vector Consistency dimension.

[0154] To learn the deviation between the user's actual behavior and the model's predicted behavior, the system defines the following loss function:

[0155] ;

[0156] In the formula, This is the user's response vector to the displayed advertisement; The embedding vector for the advertisement; For user factor vectors; For loss function, This represents a matching function for predicting user responses; in one implementation, the matching function is defined as:

[0157] ;

[0158] In the formula, The Sigmoid function is used to normalize the matching results to [0, 1]. This represents a similarity score between a user and an ad vector; This represents a matching function that predicts user responses. This is the inner product operation between the user and the ad vector.

[0159] After constructing the loss function, the system applies it to the user factor vector. Calculate the gradient to obtain the updated gradient used to correct the vector. The formula is as follows:

[0160] ;

[0161] In the formula, For the loss function on the user factor vector The gradient; loss function pairs The partial derivatives; For predicted advertising response; Activation function The derivative; this gradient is used to update A step size adjustment can be performed in the following way:

[0162] ;

[0163] In the formula, The learning rate controls the update magnitude.

[0164] In some embodiments, when the response behavior is complex, a multi-dimensional labeling model can be used to extend the loss function to a weighted multi-task loss. For example, a joint loss can be constructed by simultaneously considering click-through rate (CTR), conversion rate (CVR), and dwell time (Dwell Time).

[0165] ;

[0166] In the formula, This represents the total loss function value. For the loss function term in click-through rate prediction; For conversion rate prediction loss; This is the dwell time loss function term.

[0167] The above combination of loss functions These are empirical weighting coefficients, which can be set using cross-validation.

[0168] In the updated The vectors are synchronized to the user profile module, and the system uses them as the basis for the next round of ad scoring and strategy generation, ensuring that the expression of user preferences can be continuously corrected with actual interaction behavior and has dynamic response capability.

[0169] For step S6, in this embodiment, as advertisements are continuously delivered and user behavior evolves, relying solely on the initial tensor modeling results for strategy matching may lead to problems such as static profiles and model aging. A cloud-edge collaborative dynamic feedback mechanism is introduced, leveraging changes in user behavior to drive model retraining, thereby achieving periodic adaptive updates to the advertising strategy.

[0170] First, the updated user behavior embedding vector is analyzed locally (i.e., on the edge device) to generate a simplified representation through structural compression, which serves as the feedback information required for cloud training.

[0171] Specifically, let the user factor vector before the update be... The updated version is Then the system can calculate its vector difference as follows:

[0172] ;

[0173] In the formula, This indicates how the user's embedding vector changes after an ad response or model update.

[0174] As an alternative, to compress transmission bandwidth and improve the semantic density of the summary, the system can further extract the direction and intensity of vector changes as feature summaries:

[0175] Vector change magnitude: ;

[0176] Direction of change (can be normalized using unit vectors);

[0177] The time of change and the type of behavior are labeled.

[0178] In one possible implementation, the digest is structured as follows:

[0179] ;

[0180] In the formula, For users Summary of behavioral vector changes; For timestamps; For scene identification; This is a response type identifier.

[0181] This summary information will be uploaded to the cloud main server as feedback data for the construction of the training sample set.

[0182] After receiving summaries of user changes from multiple edge nodes, the cloud system uses these summaries to construct a fine-tuning sample set for the behavioral tensor model. These samples can be viewed as signals of implicit preference changes, used to reconstruct the latent factor parameter set. .

[0183] As an implementation method, tensor modeling adopts a CP decomposition structure, and the objective function is:

[0184] ;

[0185] In the formula, Reconstruct the loss function using tensors; A set of indices for non-zero observation data; These are the original observations in the third-order behavior tensor; Ranks for user, time, and location, respectively. Quantity; The rank of a tensor decomposition represents the number of dimensions of the latent space; This represents the Frobenius norm, used for regularization constraints; Regularity coefficient.

[0186] During fine-tuning, the system not only uses tensor reconstruction error as the objective function, but also optimizes performance metrics such as ad click-through rate (CTR) and conversion rate (CVR) in parallel, forming a multi-objective optimization structure, thereby balancing prediction accuracy and behavior matching efficiency in model updates.

[0187] After training, the system structurally encodes the latest round of latent factor parameters and distributes them to edge nodes in the form of modules. Specifically, the distribution may include:

[0188] The vector corresponding to the set of active users;

[0189] Time segment embedding matrix, geographic location embedding matrix;

[0190] Strategy adjustment parameters, such as recommendation factor weights and entropy constraint thresholds.

[0191] After receiving parameters, the edge nodes can automatically perform vector replacement or parameter interpolation operations in the strategy module to complete the distributed collaborative evolution of the local model.

[0192] In some embodiments, a version control mechanism may be introduced to improve system robustness, with each round of parameter distribution including a model number to prevent old parameters from being loaded repeatedly or causing concurrent conflicts.

[0193] In some embodiments, a version control mechanism may be introduced to improve system robustness. Each round of parameter distribution includes a model number (model_ver) to prevent old parameters from being loaded repeatedly or causing concurrent conflicts.

[0194] The advertising delivery system based on charging piles described below can be referred to in conjunction with the advertising delivery method based on charging piles described above.

[0195] Please see the appendix Figure 2 The present invention also provides an advertising delivery system based on charging piles, comprising:

[0196] The data acquisition module is used to collect various types of data, including user behavior data, device status data, scene environment data, and advertising response data.

[0197] The tensor modeling module is used to construct a third-order behavioral tensor based on the data provided by the data acquisition module.

[0198] The strategy optimization module is used to calculate the probability distribution of users accepting ads based on the generated user behavior embedding vectors, and to use information entropy to evaluate the potential effect of ad acceptance and generate the best delivery strategy.

[0199] The strategy filtering module is used to filter out the best candidate ads based on the change in information entropy, according to the best delivery strategy.

[0200] The edge delivery module is used to actually push advertisements at the charging pile and evaluate the advertising effect through the collected feedback information;

[0201] The strategy update module is used to receive feedback information from the edge delivery module, update the parameters of the delivery strategy, and synchronize the updated strategy to the devices at the charging pile.

[0202] The system in this embodiment can be used to execute the above method embodiments, and its principle and technical effect are similar, so they will not be described again here.

[0203] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for advertising placement based on charging piles, characterized in that, Includes the following steps: S1. Collect user interaction data with charging piles, user behavior data, location information, time information, and environmental status data to construct a dataset of behavioral characteristics; S2. Based on the dataset of behavioral features, construct a third-order behavioral tensor with user identifier, time segment and geographical location as dimensions, and decompose the third-order behavioral tensor to generate the corresponding user behavior embedding vector; S3. Match the user behavior embedding vector with the candidate ad set, establish a probability distribution model of the user for each candidate ad, and calculate the information entropy value of ad selection based on the probability distribution model of each candidate ad. S4. Identify and filter the target advertisement that causes the largest decrease in entropy based on the information entropy value, and use it as the current advertisement to be delivered; S5. Display the target advertisement on the charging pile terminal, collect user response behavior data to the target advertisement, and combine the display records of the delivered advertisement with user clicks, browsing time and interaction path to generate a response vector; calculate the update gradient based on the response vector and the delivered advertisement vector, which is used to correct the user factor vector in the tensor decomposition model; dynamically map the changes in user behavior preferences in the third-order tensor through the corrected user factor vector, and update the user behavior characteristics based on the response behavior. S6. Feed the updated user behavior features back to the cloud system for retraining the behavior tensor model and dynamically adjusting the advertising strategy.

2. The advertising placement method based on charging piles according to claim 1, characterized in that, In step S1, constructing the dataset of behavioral features includes the following steps: Based on the collected user-charging pile interaction data, 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 to a structured vector containing user identifier, charging duration, time tag, weather status, and scene number; The structured vector serves as the data foundation for subsequent tensor modeling steps and is input into the tensor dimension generator to define the index field of the behavioral tensor.

3. The advertising placement method 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 ID, time segment ID, and geographic region ID are used as the three dimensions of the third-order behavior tensor. During the tensor construction process, the existence of behavioral events is used as the criterion for setting tensor values, thus forming a third-order sparse behavioral tensor. The tensor is decomposed using the CP decomposition method that minimizes the reconstruction error to generate user embedding vectors.

4. The advertising placement method based on charging piles according to claim 3, characterized in that, The tensor is decomposed using the CP decomposition method that minimizes the reconstruction error. The minimum reconstruction error is achieved using the following formula: ； In the formula, Data representing user, time, and location interaction behaviors. These represent the rank-one components in the user, time, and location dimensions, respectively. The dimension of the latent space is represented. This represents the outer product of vectors.

5. The advertising placement method based on charging piles according to claim 1, characterized in that, In step S3, calculating the information entropy value of ad selection based on the probability distribution model of each candidate ad includes 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. The probability of a user selecting each advertisement is obtained by normalizing all advertisement scores using Softmax. The Softmax normalization formula is as follows: ； In the formula, This indicates the probability that a user will select the advertisement. This represents the sum of the scores for all advertisements. Indicates the first The score of each advertisement, is the base of the natural logarithm; Construct an information entropy function based on probability distribution. The information entropy is used to quantify the distribution of the current advertisements; In the formula, The information entropy chosen for advertising This indicates the probability that a user will select the advertisement. For the first The logarithm of the probability of selecting an advertisement. This represents the total number of advertisements.

6. The advertising placement method based on charging piles according to claim 1, characterized in that, In step S4, identifying and filtering the target advertisement that causes the largest decrease in entropy based on the information entropy value includes the following steps: The established probability distribution model is invoked to simulate the changes in user behavior after each candidate ad is delivered, and the predicted probability distribution is generated. For each candidate ad, calculate the change in information entropy, which is used as the change in information structure after the ad is delivered. Based on the information entropy changes of all candidate ads, the ad with the largest decrease in information entropy is selected and a constraint is set for its information entropy decrease to be met, which is then used as the final ad for delivery.

7. The advertising placement method based on charging piles according to claim 1, characterized in that, The calculation of the update gradient based on the response vector and the delivered ad vector includes the following steps: Let the user's response vector to the displayed advertisement be set as... The embedding vector of the advertisement is The current user factor vector is ; The loss function for calculating user behavior preference error is constructed as follows: ； In the formula, For loss function, This represents a matching function for predicting user responses; By calculating the loss function with respect to Take the derivative to obtain the update gradient of the user factor vector: ； In the formula, For the loss function on the user factor vector gradient, loss function pairs The partial derivatives, For predicted advertising response, Activation function The derivative of .

8. The advertising placement method based on charging piles according to claim 1, characterized in that, In step S6, retraining the behavioral tensor model and dynamically adjusting the advertising delivery strategy includes the following steps: The updated user embedding vector changes are summarized, and feature data of the vector change direction and magnitude are extracted and uploaded to the cloud. The cloud system uses change summaries uploaded from multiple edge nodes to construct a sample set for fine-tuning the tensor model, and then retrains the latent factor parameter set based on this sample set. After training is complete, the updated parameters are sent to edge devices to complete the distributed collaborative evolution of the policy model.

9. An advertising delivery system based on charging piles, applied to the advertising delivery method based on charging piles as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect various types of data, including user behavior data, device status data, scene environment data, and advertising response data. The tensor modeling module is used to construct a third-order behavioral 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 vectors, and to use information entropy to evaluate the potential effect of ad acceptance and generate the best delivery strategy. The strategy filtering module is used to filter out the best candidate ads based on the change in information entropy, according to the best delivery strategy. The edge delivery module is used to actually push advertisements at the charging pile end, collect user response behavior data to target advertisements, and combine the display records of delivered advertisements with user clicks, browsing time and interaction path to generate response vectors; calculate update gradients based on response vectors and delivered advertisement vectors, which are used to back-correct user factor vectors in tensor decomposition model; through the corrected user factor vectors, dynamically map changes in user behavior preferences in the third-order tensor, thereby updating user behavior characteristics based on response behavior; The strategy update module is used to receive feedback information from the edge delivery module, update the parameters of the delivery strategy, and synchronize the updated strategy to the devices at the charging pile.

Images