Payment scene advertisement recommendation method and system based on user consumption behavior portrait

Abstract

This invention relates to the field of advertising recommendation technology, and in particular to a method and system for advertising recommendation in payment scenarios based on user consumption behavior profiles. The method includes: acquiring multimodal consumption behavior data and payment scenario information, preprocessing and encoding them into context vectors; extracting payment semantic feature vectors and video behavior feature vectors, and constructing user consumption behavior profile vectors through graph attention network association propagation; constructing a payment intent prediction model based on a Transformer teacher network and a spatiotemporal graph attention student network, inputting the profile vector and context vectors to obtain payment intent prediction results and advertising matching scores, and generating a first advertising recommendation strategy; real-time collection and updating of features and calculation of feature drift, optimizing the first strategy to obtain a second advertising recommendation strategy; and executing advertising delivery according to the second strategy. This invention achieves high-precision, low-latency, and highly adaptable advertising recommendation in payment scenarios.

Description

Technical Field

[0001] This invention relates to the field of advertising recommendation technology, and in particular to a method and system for advertising recommendation in payment scenarios based on user consumption behavior profiles. Background Technology

[0002] With the widespread adoption of mobile payments and smart devices, payment scenarios have become a crucial entry point for precise advertising. Existing advertising recommendation technologies primarily rely on offline modeling based on users' historical click records and static profiles. However, in payment scenarios, these technologies suffer from insufficient real-time performance and weak context awareness, making it difficult to capture the dynamic intent of users at the moment of payment.

[0003] Furthermore, existing technologies do not fully utilize multimodal consumer behavior data in payment scenarios, failing to effectively integrate payment text information with consumer behavior video information. Moreover, large-scale deep learning models have high computational complexity, making them difficult to deploy efficiently in environments with limited payment terminal resources. At the same time, existing recommendation strategies lack dynamic optimization mechanisms based on real-time behavioral changes, resulting in a time lag between ad recommendations and users' current consumption intentions, leading to low recommendation conversion rates. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for advertising recommendation in payment scenarios based on user consumption behavior profiles. The aim is to achieve high-precision, low-latency, and highly adaptable advertising recommendations in payment scenarios by deeply integrating multimodal consumption behavior data and constructing real-time profiles, combined with a lightweight payment intent prediction model and a feature drift-driven strategy optimization mechanism.

[0005] To achieve the above objectives, this invention provides a method for recommending advertisements in payment scenarios based on user consumption behavior profiles, comprising the following steps: S1. Obtain multimodal consumption behavior data and payment scenario information of users in payment scenarios, perform preprocessing operations such as timestamp alignment, missing value imputation and outlier removal on the multimodal consumption behavior data, generate a standardized consumption behavior dataset, and encode the payment scenario information into a payment scenario context vector; S2. Based on the standardized consumer behavior dataset, obtain the payment semantic feature vector and the video behavior feature vector; apply a graph attention network to perform node association and feature propagation on the payment semantic feature vector and the video behavior feature vector to construct a user consumer behavior profile vector; S3. A payment intention prediction model is constructed using a teacher network based on the Transformer architecture and a student network based on a spatiotemporal graph attention network. The user consumption behavior profile vector and the payment scenario context vector are input into the payment intention prediction model to obtain the payment intention prediction result and the ad matching score output by the model. The first ad recommendation strategy is generated based on the payment intention prediction result and the ad matching score. S4. Collect payment behavior update features and video behavior update features in real time, calculate feature drift, and optimize the first advertising recommendation strategy based on feature drift to obtain the second advertising recommendation strategy. S5. Execute ad delivery based on the second ad recommendation strategy.

[0006] Preferably, in S1, the payment scenario information includes the payment terminal type, geographical location, network status, current time period, and merchant category code; the preprocessing operations for multimodal consumption behavior data include: establishing a unified time benchmark based on the timestamp field and resampling, performing linear interpolation on missing segments, and removing outliers that exceed three times the standard deviation of the mean.

[0007] Preferably, in S2, the payment semantic feature vector and video behavior feature vector obtained based on the standardized consumer behavior dataset include: A pre-trained language model is used to semantically encode transaction notes and merchant names to generate payment semantic feature vectors; a convolutional neural network is used to extract frame-level features from consumer behavior videos, and a spatiotemporal convolutional network is used for temporal aggregation to generate video behavior feature vectors.

[0008] Preferably, in S2, the node feature update formula for the graph attention network is: ; Attention Score: ; in, Let i be the feature vector of node i in the l-th layer. For learnable weight matrix, For attention vectors, For splicing operations, Let be the set of neighbors of node i.

[0009] Preferably, in S3, the ad matching score The calculation formula is: ; in, C is a user consumption behavior profile vector, and C is a payment scenario context vector. For payment intention prediction function, For context enhancement functions, , All are weighting coefficients, b is the bias term, and σ is the activation function; the payment intention prediction result is determined by comparing the score with the preset score threshold.

[0010] Preferably, in S4, the formula for calculating the feature drift is: ; in, , These are the real-time payment behavior update features and the historical payment behavior features, respectively. , These are respectively real-time video behavior update features and historical video behavior features. , All are weighting coefficients. Represents the L2 norm; The first ad recommendation strategy is optimized based on the feature drift, including determining whether the feature drift exceeds a preset drift threshold. If it does, the ad candidate set is replaced and the delivery parameters are adjusted.

[0011] Preferably, in S4, optimizing the first ad recommendation strategy based on feature drift includes: The feature drift is input into the reinforcement learning model. The click feedback probability and conversion probability of the current ad candidate set are used as reward signals. The policy network generates ad position adjustment actions, performs real-time reordering and deduplication on the ad candidate set, and adjusts the display position and delivery timing in the first ad recommendation strategy to obtain the second ad recommendation strategy.

[0012] Preferred payment scenario advertising recommendation systems based on user consumption behavior profiles include: The data acquisition module is used to acquire multimodal consumption behavior data and payment scenario information of users in payment scenarios. It performs preprocessing operations such as timestamp alignment, missing value imputation and outlier removal on the multimodal consumption behavior data to generate a standardized consumption behavior dataset and encodes the payment scenario information into a payment scenario context vector. The vector construction module is used to obtain payment semantic feature vectors and video behavior feature vectors based on a standardized consumer behavior dataset; a graph attention network is applied to perform node association and feature propagation on the payment semantic feature vectors and video behavior feature vectors to construct user consumer behavior profile vectors. The strategy generation module is used to build a payment intention prediction model based on the teacher network based on the Transformer architecture and the student network based on the spatiotemporal graph attention network. The user consumption behavior profile vector and the payment scenario context vector are input into the payment intention prediction model to obtain the payment intention prediction result and the ad matching score output by the model. The first ad recommendation strategy is generated based on the payment intention prediction result and the ad matching score. The strategy optimization module is used to collect payment behavior update features and video behavior update features in real time, calculate feature drift, and optimize the first advertising recommendation strategy based on the feature drift to obtain the second advertising recommendation strategy. The ad delivery module is used to execute ad delivery based on the second ad recommendation strategy.

[0013] The advantages and beneficial effects of this invention compared to the prior art are: 1. This invention utilizes the collaborative processing of a pre-trained language model, a convolutional neural network, and a graph attention network to perform cross-modal association and feature propagation of payment semantic feature vectors and video behavior feature vectors, thereby constructing an accurate user consumption behavior profile vector and improving the accuracy of user intent recognition in payment scenarios.

[0014] 2. This invention uses a teacher network based on the Transformer architecture and a student network based on a spatiotemporal graph attention network to construct a payment intention prediction model. Knowledge transfer is achieved through the attention distillation loss function, which significantly reduces the number of model parameters and inference latency while ensuring cross-modal understanding capabilities, and is suitable for the resource-constrained environment of payment terminals.

[0015] 3. This invention inputs the payment scenario context vector and the user consumption behavior profile vector into the payment intent prediction model, so that the ad matching score fully integrates the user's historical consumption attributes and current payment scenario information, thereby enhancing the fit between ad recommendations and real-time payment context.

[0016] 4. This invention dynamically optimizes the first advertising recommendation strategy by collecting payment behavior update features and video behavior update features in real time and calculating feature drift, thereby obtaining a second advertising recommendation strategy. This shortens the response time lag between advertising strategies and changes in user consumption intentions, and improves recommendation timeliness and conversion rate.

[0017] 5. This invention performs real-time reordering and deduplication on the candidate ad set based on feature drift, optimizing ad display position while controlling duplicate exposure, thus balancing recommendation accuracy and user experience.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart of the payment scenario advertising recommendation method based on user consumption behavior profiles in an embodiment of the present invention; Figure 2 This is a structural diagram of a payment scenario advertising recommendation system based on user consumption behavior profiles, according to an embodiment of the present invention. Detailed Implementation

[0020] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this invention provides a method for recommending advertisements in payment scenarios based on user consumption behavior profiles, including the following steps: S1. Obtain multimodal consumption behavior data and payment scenario information of users in payment scenarios, perform preprocessing operations such as timestamp alignment, missing value imputation and outlier removal on the multimodal consumption behavior data, generate a standardized consumption behavior dataset, and encode the payment scenario information into a payment scenario context vector; S2. Based on the standardized consumer behavior dataset, obtain the payment semantic feature vector and the video behavior feature vector; apply a graph attention network to perform node association and feature propagation on the payment semantic feature vector and the video behavior feature vector to construct a user consumer behavior profile vector; S3. A payment intention prediction model is constructed using a teacher network based on the Transformer architecture and a student network based on a spatiotemporal graph attention network. The user consumption behavior profile vector and the payment scenario context vector are input into the payment intention prediction model to obtain the payment intention prediction result and the ad matching score output by the model. The first ad recommendation strategy is generated based on the payment intention prediction result and the ad matching score. S4. Collect payment behavior update features and video behavior update features in real time, calculate feature drift, and optimize the first advertising recommendation strategy based on feature drift to obtain the second advertising recommendation strategy. S5. Execute ad delivery based on the second ad recommendation strategy.

[0023] Preferably, in S1, the payment scenario information includes the payment terminal type, geographical location, network status, current time period, and merchant category code; the preprocessing operations for multimodal consumption behavior data include: establishing a unified time benchmark based on the timestamp field and resampling, performing linear interpolation on missing segments, and removing outliers that exceed three times the standard deviation of the mean.

[0024] In one embodiment, the "Merchant Category Code" (MCC) in the payment scenario information is a four-digit code assigned to merchants by the card organization to identify the merchant's industry category; for example, 5411 represents a supermarket and 5812 represents a restaurant. The preprocessing stage establishes a unified time benchmark to ensure that multi-source data is aligned in the time dimension; linear imputation fills in missing values ​​based on adjacent valid data points; outlier removal uses the 3σ criterion to exclude outliers caused by network jitter or sensor malfunctions.

[0025] Preferably, in S2, the payment semantic feature vector and video behavior feature vector obtained based on the standardized consumer behavior dataset include: A pre-trained language model is used to semantically encode transaction notes and merchant names to generate payment semantic feature vectors; a convolutional neural network is used to extract frame-level features from consumer behavior videos, and a spatiotemporal convolutional network is used for temporal aggregation to generate video behavior feature vectors.

[0026] In one embodiment, a pre-trained language model (such as a BERT-like model based on bidirectional encoder representation) acquires semantic understanding capabilities through pre-training on a large-scale corpus, mapping transaction notes and merchant names into high-dimensional dense vectors. A convolutional neural network (CNN) extracts local visual features from video frames of consumer behavior by sliding convolutional kernels; a spatiotemporal convolutional network (such as 3D-CNN or (2+1)D-CNN) aggregates these frame-level features along the temporal dimension, capturing the sequential patterns of continuous actions taken by users while browsing products, and generating video behavior feature vectors.

[0027] Preferably, in S2, the node feature update formula for the graph attention network is: ; Attention Score: ; in, Let i be the feature vector of node i in the l-th layer. For learnable weight matrix, For attention vectors, For splicing operations, Let be the set of neighbors of node i.

[0028] In one embodiment, during the node feature update process, each node aggregates the feature information of its neighborhood set (i.e., other nodes that have consumption associations with the node, such as different transactions or different browsing records of the same type of goods under the same merchant); the attention coefficient calculates the association strength between nodes through a learnable weight matrix and attention vector, the LeakyReLU activation function introduces nonlinearity, and softmax normalization ensures that the weight sum is 1; the concatenation operation connects the transformed features of the source node and the neighboring nodes along the channel dimension to enhance feature discriminability.

[0029] Preferably, in S3, the ad matching score The calculation formula is: ; in, C is a user consumption behavior profile vector, and C is a payment scenario context vector. For payment intention prediction function, For context enhancement functions, , All are weighting coefficients, b is the bias term, and σ is the activation function; the payment intention prediction result is determined by comparing the score with the preset score threshold.

[0030] Specifically, the preset score threshold is a payment intent discrimination threshold, used to discretize continuous ad matching scores into binary or multi-category intent categories, thereby triggering ad recommendation actions of varying intensities. This threshold is determined during the offline training phase of the model by analyzing the distribution patterns of ad matching scores in historical samples and their correlation with corresponding actual conversion results. Specifically, it can be calibrated using the principle of maximizing the area under the receiver operating characteristic curve or the balance point of the precision-recall curve. Alternatively, the threshold can be dynamically fine-tuned according to business operation needs to achieve a balance between recommendation accuracy and user interruption frequency.

[0031] In one embodiment, the payment intent prediction function represents the mapping relationship between user consumption profile and payment intent, and the context enhancement function highlights the impact of the current payment scenario on intent judgment; the two are weighted and fused, and then mapped to a score value in the range of 0 to 1 by an activation function. The payment intent prediction result is determined by comparing the score with a preset threshold: a score higher than the threshold is judged as high payment intent, triggering precise ad push; a score lower than the threshold is judged as low intent, reducing the frequency of ad interruptions.

[0032] Preferably, in S4, the formula for calculating the feature drift is: ; in, , These are the real-time payment behavior update features and the historical payment behavior features, respectively. , These are respectively real-time video behavior update features and historical video behavior features. , All are weighting coefficients. Represents the L2 norm; The first ad recommendation strategy is optimized based on the feature drift, including determining whether the feature drift exceeds a preset drift threshold. If it does, the ad candidate set is replaced and the delivery parameters are adjusted.

[0033] Specifically, the preset drift threshold is a feature drift trigger threshold used to determine whether the deviation of a user's real-time consumption behavior from the historical baseline has reached the critical condition requiring an update to the recommendation strategy. This threshold is determined during the system initialization phase by analyzing the statistical distribution of historical feature drift for the user group. Specifically, it is selected from the upper quartile of the drift distribution for all users, or based on the quartile-drift sensitivity curve, to achieve a balance between strategy update frequency and computational resource consumption. Alternatively, it can be adaptively segmented based on different merchant categories or time periods to avoid excessive strategy updates caused by normal consumption fluctuations.

[0034] In one embodiment, feature drift is measured by the L2 norm (i.e., Euclidean distance, calculated by taking the square root of the sum of the squares of the differences between corresponding elements of two vectors) to determine the spatial deviation between real-time and historical features; weighting coefficients balance the contribution of payment behavior and video behavior to drift. During optimization, if the drift exceeds a preset threshold, it indicates a significant change in user consumption interests. The system then replaces the ad candidate set and adjusts the delivery parameters to ensure that recommended content is updated synchronously with the user's status.

[0035] Preferably, in S4, optimizing the first ad recommendation strategy based on feature drift includes: The feature drift is input into the reinforcement learning model. The click feedback probability and conversion probability of the current ad candidate set are used as reward signals. The policy network generates ad position adjustment actions, performs real-time reordering and deduplication on the ad candidate set, and adjusts the display position and delivery timing in the first ad recommendation strategy to obtain the second ad recommendation strategy.

[0036] In one embodiment, the reinforcement learning model uses feature drift as the input to the environment state and the click feedback probability and conversion probability of the current ad candidate set as the reward signal to measure the positive user feedback brought by the recommendation action; the policy network (such as a neural network based on the policy gradient method) outputs ad position adjustment actions, including improving the ranking of highly relevant ads, removing duplicate exposure items, and adjusting the timing of delivery, so as to realize the real-time reordering and deduplication of the recommendation list.

[0037] Preferred, such as Figure 2 As shown, the payment scenario advertising recommendation system based on user consumption behavior profiles includes: The data acquisition module is used to acquire multimodal consumption behavior data and payment scenario information of users in payment scenarios. It performs preprocessing operations such as timestamp alignment, missing value imputation and outlier removal on the multimodal consumption behavior data to generate a standardized consumption behavior dataset and encodes the payment scenario information into a payment scenario context vector. The vector construction module is used to obtain payment semantic feature vectors and video behavior feature vectors based on a standardized consumer behavior dataset; a graph attention network is applied to perform node association and feature propagation on the payment semantic feature vectors and video behavior feature vectors to construct user consumer behavior profile vectors. The strategy generation module is used to build a payment intention prediction model based on the teacher network based on the Transformer architecture and the student network based on the spatiotemporal graph attention network. The user consumption behavior profile vector and the payment scenario context vector are input into the payment intention prediction model to obtain the payment intention prediction result and the ad matching score output by the model. The first ad recommendation strategy is generated based on the payment intention prediction result and the ad matching score. The strategy optimization module is used to collect payment behavior update features and video behavior update features in real time, calculate feature drift, and optimize the first advertising recommendation strategy based on the feature drift to obtain the second advertising recommendation strategy. The ad delivery module is used to execute ad delivery based on the second ad recommendation strategy.

[0038] In one embodiment, the data acquisition module is deployed at the edge of the payment terminal, responsible for raw data collection and cleaning; the vector construction module is deployed on the edge computing node, completing feature extraction and profile generation; the strategy generation module is deployed on a lightweight inference engine, performing payment intent prediction; the strategy optimization module runs in a streaming computing mode, monitoring feature drift in real time and triggering strategy updates; and the advertising delivery module connects to the payment interface SDK, completing ad rendering according to the second strategy. Each module's function and methodological steps correspond one-to-one, supporting distributed collaborative operation.

[0039] The following verification is based on a specific embodiment.

[0040] Take a user's QR code payment scenario in a large chain supermarket as an example. When a user completes a transaction using a mobile payment terminal, the system executes the following process: S1: Collect the user's consumption records (transaction notes, merchant name, consumption amount) at the supermarket over the past 30 days, as well as the current payment scenario information (payment terminal type: smartphone, geographical location: a supermarket in Chaoyang District, Beijing, network status: 5G, current time: Saturday 18:30, merchant category code: 5411). Perform timestamp alignment to second-level precision on the multimodal consumption behavior data. Perform linear imputation on 3 missing amount records due to network interruption. Remove 1 abnormally large amount record (exceeding 3 times the standard deviation of the mean) caused by sensor mis-touch. Generate a standardized consumption behavior dataset. Encode the payment scenario information into a 256-dimensional payment scenario context vector.

[0041] S2: The pre-trained language model BERT-base is used to semantically encode the transaction notes "purchase of fresh vegetables" and the merchant name "XX Supermarket," generating a 768-dimensional payment semantic feature vector. A ResNet-50 convolutional neural network is used to extract frame-level features from a 15-second surveillance video of the user browsing the supermarket's merchandise area, extracting 2048-dimensional features per frame. These features are then aggregated along the time dimension using a 3D-CNN to form a 512-dimensional video behavior feature vector. A two-layer graph attention network is used to associate the two vectors: the number of attention heads is set to 8, and the neighborhood set is taken from transaction nodes of similar merchants within the same user over the past 7 days. After feature propagation, a 1024-dimensional user consumption behavior profile vector is generated.

[0042] S3: Construct a payment intent prediction model, where the teacher network adopts a 6-layer Transformer architecture (512 hidden layer dimensions, 8 attention heads), and the student network adopts a 4-layer spatiotemporal graph attention network (256 hidden layer dimensions, 4 attention heads). The user consumption behavior profile vector and the payment scenario context vector are concatenated and input into the student network. After knowledge distillation training, the model outputs a payment intent prediction result of "high intent" (Score=0.87, preset threshold 0.75), with an ad matching score of 0.87. The first ad recommendation strategy is generated: the candidate set includes fresh food coupons, daily necessities discounts, and recommendations for nearby restaurants, sorted in descending order of score.

[0043] S4: Real-time collection of payment behavior update features within 5 seconds after user payment completion (payment amount updated from the historical average of 120 yuan to 280 yuan this time) and video behavior update features (user lingers in the prepared food area for 3 seconds after payment). Calculation of feature drift: Payment behavior drift. =160, Video Behavior Drift =45, taking β1=0.6 and β2=0.4, we get ΔD=114. Since ΔD exceeds the preset threshold of 80, strategy optimization is triggered: the feature drift is input into the reinforcement learning model based on the PPO algorithm, with a click feedback probability of 0.12 and a conversion probability of 0.08 as the reward signal. The strategy network outputs the following actions: the "Delicious Food Section Discount Coupon" is moved to the top position, the previously exposed "Fresh Food Coupon" is removed, and the timing of the ad placement is adjusted to a pop-up on the payment completion page. A second ad recommendation strategy is generated.

[0044] S5: The payment completion page displays the "Delicious Food Section Discount Coupon" first, "Daily Necessities Discount" second, and "Nearby Restaurant Recommendations" last according to the second strategy. Users click on "Delicious Food Section Discount Coupon" and claim it, thus realizing advertising.

[0045] A / B testing has verified that, compared with traditional static recommendations, the system using this embodiment has increased ad click-through rate by 23.6%, conversion rate by 17.2%, and reduced average inference latency from 420ms to 85ms, meeting the low latency requirements of payment scenarios.

[0046] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for recommending advertisements in payment scenarios based on user consumption behavior profiles, characterized in that, Includes the following steps: S1. Obtain multimodal consumption behavior data and payment scenario information of users in payment scenarios, perform preprocessing operations such as timestamp alignment, missing value imputation and outlier removal on the multimodal consumption behavior data, generate a standardized consumption behavior dataset, and encode the payment scenario information into a payment scenario context vector; S2. Based on the standardized consumer behavior dataset, obtain the payment semantic feature vector and the video behavior feature vector; apply a graph attention network to perform node association and feature propagation on the payment semantic feature vector and the video behavior feature vector to construct a user consumer behavior profile vector; S3. A payment intention prediction model is constructed using a teacher network based on the Transformer architecture and a student network based on a spatiotemporal graph attention network. The user consumption behavior profile vector and the payment scenario context vector are input into the payment intention prediction model to obtain the payment intention prediction result and the ad matching score output by the model. The first ad recommendation strategy is generated based on the payment intention prediction result and the ad matching score. S4. Collect payment behavior update features and video behavior update features in real time, calculate feature drift, and optimize the first advertising recommendation strategy based on feature drift to obtain the second advertising recommendation strategy. S5. Execute ad delivery based on the second ad recommendation strategy.

2. The method for recommending advertisements in payment scenarios based on user consumption behavior profiles according to claim 1, characterized in that, In S1, the payment scenario information includes the payment terminal type, geographical location, network status, current time period, and merchant category code; Preprocessing operations for multimodal consumer behavior data include: establishing a unified time base based on the timestamp field and resampling, performing linear interpolation on missing segments, and removing outliers that exceed three times the standard deviation of the mean.

3. The method for recommending advertisements in payment scenarios based on user consumption behavior profiles according to claim 2, characterized in that, In S2, the payment semantic feature vector and video behavior feature vector obtained based on the standardized consumer behavior dataset include: A pre-trained language model is used to semantically encode transaction notes and merchant names to generate payment semantic feature vectors; a convolutional neural network is used to extract frame-level features from consumer behavior videos, and a spatiotemporal convolutional network is used for temporal aggregation to generate video behavior feature vectors.

4. The method for recommending advertisements in payment scenarios based on user consumption behavior profiles according to claim 3, characterized in that, In S2, the node feature update formula for the graph attention network is: ； Attention Score: ； in, Let i be the feature vector of node i in the l-th layer. For learnable weight matrix, For attention vectors, For splicing operations, Let be the set of neighbors of node i.

5. The method for recommending advertisements in payment scenarios based on user consumption behavior profiles according to claim 4, characterized in that, In S3, ad match score The calculation formula is: ； in, C is a user consumption behavior profile vector, and C is a payment scenario context vector. For payment intention prediction function, For context enhancement functions, , All are weighting coefficients, b is the bias term, and σ is the activation function; the payment intention prediction result is determined by comparing the score with the preset score threshold.

6. The method for recommending advertisements in payment scenarios based on user consumption behavior profiles according to claim 5, characterized in that, In S4, the formula for calculating the feature drift is: ； in, , These are the real-time payment behavior update features and the historical payment behavior features, respectively. , These are respectively real-time video behavior update features and historical video behavior features. , All are weighting coefficients. Represents the L2 norm; The first ad recommendation strategy is optimized based on the feature drift, including determining whether the feature drift exceeds a preset drift threshold. If it does, the ad candidate set is replaced and the delivery parameters are adjusted.

7. The method for recommending advertisements in payment scenarios based on user consumption behavior profiles according to claim 6, characterized in that, In S4, the optimization of the first advertisement recommendation strategy based on feature drift includes: The feature drift is input into the reinforcement learning model. The click feedback probability and conversion probability of the current ad candidate set are used as reward signals. The policy network generates ad position adjustment actions, performs real-time reordering and deduplication on the ad candidate set, and adjusts the display position and delivery timing in the first ad recommendation strategy to obtain the second ad recommendation strategy.

8. A payment scenario advertising recommendation system based on user consumption behavior profiles, characterized in that, include: The data acquisition module is used to acquire multimodal consumption behavior data and payment scenario information of users in payment scenarios. It performs preprocessing operations such as timestamp alignment, missing value imputation and outlier removal on the multimodal consumption behavior data to generate a standardized consumption behavior dataset and encodes the payment scenario information into a payment scenario context vector. The vector construction module is used to obtain payment semantic feature vectors and video behavior feature vectors based on a standardized consumer behavior dataset; a graph attention network is applied to perform node association and feature propagation on the payment semantic feature vectors and video behavior feature vectors to construct user consumer behavior profile vectors. The strategy generation module is used to build a payment intention prediction model based on the teacher network based on the Transformer architecture and the student network based on the spatiotemporal graph attention network. The user consumption behavior profile vector and the payment scenario context vector are input into the payment intention prediction model to obtain the payment intention prediction result and the ad matching score output by the model. The first ad recommendation strategy is generated based on the payment intention prediction result and the ad matching score. The strategy optimization module is used to collect payment behavior update features and video behavior update features in real time, calculate feature drift, and optimize the first advertising recommendation strategy based on the feature drift to obtain the second advertising recommendation strategy. The ad delivery module is used to execute ad delivery based on the second ad recommendation strategy.

Images