A method and system for intelligent delivery of mobile banking banner ads based on a fusion model

Abstract

This invention provides a method and system for intelligent banner ad delivery in mobile banking based on a fusion model. The method includes: acquiring user offline features, user dynamic features, and ad feature sets; inputting user offline features and ad features into a first model to generate a first predicted score set, and inputting user dynamic features and ad features into a second model to generate a second predicted score set; weighting and fusing the two score sets according to user type to obtain a fused predicted score set; querying a compliance control matrix based on user risk level and ad risk level to obtain a compliance adjustment coefficient, and performing non-linear adjustments on the fused scores to generate a final ad score set; and selecting the ad with the highest score for placement on the banner. The system can be deployed on a server cluster with heterogeneous model processing capabilities, achieving integrated processing of heterogeneous modeling of static and dynamic features, prediction fusion, and compliance control, thereby improving the accuracy, real-time performance, and compliance of ad recommendations.

Description

Technical Field

[0001] This invention relates to the field of financial technology, specifically to a method and system for intelligent delivery of mobile banking banner ads based on a fusion model. Background Technology

[0002] With the development of mobile internet, mobile banking has become an important channel for financial services. Banner ads, as a core exposure area on mobile apps, are of significant value in improving customer activity and cross-selling conversions. To achieve precise ad targeting and intelligent recommendations, the ad delivery system in bank apps has undergone a three-stage technological evolution: from static configuration to machine learning-driven approaches. The first stage, rule-based static delivery systems, relied primarily on basic customer attributes (such as age, gender, and asset level) for simple matching, failing to reflect individual user differences and content suitability. The second stage introduced collaborative filtering algorithms, performing nearest-neighbor recommendations based on user group behavior, providing preliminary personalization capabilities, but still struggling to handle cold-start users and ad diversity. The third stage began applying machine learning models (such as logistic regression (LR) and gradient boosting tree (GBDT)) for ad click-through rate prediction, extracting multi-dimensional attributes through feature engineering, thus improving prediction accuracy and delivery effectiveness.

[0003] However, existing technologies still face the following significant shortcomings in mobile banking banner ad placement scenarios: 1) Poor ad type adaptability: Existing systems generally treat candidate ads as homogeneous content, failing to fully identify the risk characteristics and reach strategy differences of product, activity, and benefit ads. In particular, product ads have serious deficiencies in risk matching, which can easily lead to compliance risks; 2) Weak compliance risk control: Current ad recommendation processes mostly adopt "post-processing" filtering, lacking risk constraint mechanisms embedded in the prediction process. At the same time, compliance assessment factors are singular, failing to combine user risk levels with ad risk levels for joint modeling; 3) Poor response timeliness: Existing recommendation systems have long update cycles, failing to effectively capture users' real-time behavioral changes, resulting in a lag in responding to short-term behavioral interests, making it difficult to support accurate switching of real-time banner ads, and affecting click-through conversion efficiency.

[0004] In summary, there is an urgent need for an intelligent advertising delivery method and system that integrates static and dynamic behavioral characteristics, supports multi-model prediction fusion, and has the ability to dynamically adjust compliance risks, so as to improve the accuracy, compliance, and real-time performance of recommendations. Summary of the Invention

[0005] This disclosure provides a method and system for intelligent delivery of mobile banking banner ads based on a fusion model.

[0006] Firstly, this disclosure provides a method for intelligent delivery of mobile banking banner ads based on a fusion model, including:

[0007] Obtain a set of user offline characteristics, user dynamic characteristics, and advertising characteristics;

[0008] The user offline features and the set of advertising features are combined one by one and input into the pre-trained first model to generate the first set of prediction scores.

[0009] The user dynamic features and the set of advertising features are combined one by one and input into the pre-trained second model to generate a second set of prediction scores.

[0010] The first predicted score set and the second predicted score set are weighted and fused to obtain a fused predicted score set.

[0011] The final advertising score set is obtained by combining the compliance adjustment coefficient with the fusion prediction score set.

[0012] Select at least one advertisement with the highest score from the final ad score set and place it in the mobile banking banner position.

[0013] Optionally, the first model is a structured gradient enhancement prediction model, used to receive a joint input vector composed of the user's static features and the advertising features, and output a first prediction score;

[0014] The second model is an embedded behavior modeling deep prediction model, which is used to receive a joint input vector composed of the user's dynamic behavior features and advertising features, and output a second prediction score.

[0015] The first prediction score and the second prediction score reflect the user's long-term static preferences and short-term behavioral interests, respectively.

[0016] Optionally, the step of weightedly fusing the first predicted score set and the second predicted score set to obtain a fused predicted score set further includes:

[0017] The current users are categorized into at least three types: new users, active users, and low-frequency users.

[0018] Different weighting coefficients are set for the first predicted score set and the second predicted score set according to different user types;

[0019] The first predicted score set and the second predicted score set are weighted and fused according to the weighting coefficients to obtain the fused predicted score set.

[0020] Optionally, obtaining the final advertising score set based on the compliance adjustment coefficient and the fused prediction score set further includes:

[0021] Based on the current user's risk level and the risk type of the candidate advertisement, query the preset compliance control matrix to obtain the compliance adjustment coefficient corresponding to each fusion prediction score;

[0022] For each score value in the fusion prediction score set, the compliance adjustment coefficient is applied to perform a nonlinear calculation to generate the final advertising score set.

[0023] Optionally, the pre-trained first model and the pre-trained second model are configured with update strategies, including:

[0024] The first model performs retraining on the full dataset daily and replaces the model on a daily schedule.

[0025] The second model uses an incremental update mechanism, acquiring new user behavior data every hour to fine-tune the model;

[0026] During the update process of the second model, the cross-feature parameters in the FM layer are iteratively updated using the stochastic gradient descent method, the embedding layer and intermediate hidden layer of the DNN layer are frozen, and only the last two layers are fine-tuned.

[0027] Optionally, the construction of the compliance control matrix includes:

[0028] Multiple user risk levels are preset, including at least conservative, moderate, and aggressive risk levels;

[0029] Configure multiple risk tags for candidate ads, including at least product category, event category, and benefit category, and their corresponding risk levels;

[0030] The user risk level and the advertising risk level are combined in pairs to construct a matrix coordinate space, and a corresponding compliance adjustment coefficient is configured for each combination. The compliance adjustment coefficient is a floating-point number between 0 and 1.

[0031] Secondly, this disclosure provides a mobile banking banner ad intelligent delivery system based on a fusion model, including:

[0032] The feature acquisition module is used to acquire user offline features, user dynamic features, and advertising feature sets;

[0033] The first scoring module is used to combine the user's offline features with the set of advertising features one by one and input them into the pre-trained first model to generate a first set of predicted scores.

[0034] The second scoring module is used to combine the user dynamic features and the set of advertising features one by one and input them into the pre-trained second model to generate a second set of predicted scores.

[0035] The fusion score module is used to perform weighted fusion of the first predicted score set and the second predicted score set to obtain a fused predicted score set;

[0036] The final scoring module is used to obtain the final advertising score set based on the compliance adjustment coefficient and the fusion prediction score set;

[0037] The ad placement module is used to select at least one ad with the highest score from the final ad score set and place it in the mobile banking banner position.

[0038] Optionally, the system is deployed on a server cluster with heterogeneous model processing capabilities, the server cluster including general-purpose CPU nodes for performing the first model prediction and GPU nodes for performing the second model.

[0039] Thirdly, this disclosure provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0040] The memory stores the instructions that the computer executes;

[0041] The processor executes computer-executable instructions stored in memory to implement the methods of this disclosure.

[0042] Fourthly, this disclosure provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods of this disclosure.

[0043] The beneficial effects of this disclosure are that, compared with the prior art, this disclosure has the following advantages:

[0044] 1) This invention introduces a dual-model structure to process users' static features and dynamic behavioral information separately. Static features are modeled using the XGBoost model to model long-term user preferences, while dynamic features are captured using the DeepFM model to capture real-time interests. This effectively integrates multi-dimensional user profiles, significantly improving the accuracy and coverage of click-through rate prediction, and meeting the high-accuracy delivery requirements of financial advertising.

[0045] 2) This invention introduces a user type classification mechanism (such as new users, active users, and low-frequency users) in the prediction score fusion stage, sets differentiated weighting coefficients based on user group characteristics, realizes a more personalized recommendation strategy, improves the matching effect of advertisements on different customer groups, and avoids the adaptation imbalance caused by a one-size-fits-all model.

[0046] 3) This invention constructs a compliance control matrix based on user risk level and advertising risk type. It intervenes in the prediction score in real time through nonlinear compliance adjustment coefficient, avoids high-risk customers from receiving high-risk advertisements, effectively avoids compliance risks, realizes the intelligent recommendation strategy of "prediction equals compliance", and improves the regulatory compliance and customer trust of the financial advertising system. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0048] Figure 1 A schematic diagram of a mobile banking banner ad intelligent delivery method based on a fusion model provided in this embodiment of the disclosure;

[0049] Figure 2 This is a schematic diagram of a fusion prediction score calculation method provided in an embodiment of the present disclosure;

[0050] Figure 3 This is a schematic diagram illustrating a final advertising score calculation method provided in an embodiment of the present disclosure;

[0051] Figure 4 This is a schematic diagram of a compliance control matrix construction method provided in an embodiment of this disclosure;

[0052] Figure 5 A schematic diagram of a mobile banking banner ad intelligent delivery system based on a fusion model provided in this embodiment of the disclosure;

[0053] Figure 6 This is a schematic diagram of a dynamic behavior feature acquisition, processing, and handling process provided in an embodiment of this disclosure.

[0054] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0055] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present disclosure more clearly, and should not be used to limit the scope of protection of the present disclosure.

[0056] Figure 1 This diagram illustrates a method for intelligent delivery of mobile banking banner ads based on a fusion model, as provided in this embodiment of the disclosure. See also... Figure 1 The following is a detailed discussion of each step in conjunction with this embodiment.

[0057] S100: Obtain user offline characteristics, user dynamic characteristics, and advertising characteristic sets.

[0058] User offline characteristics mainly include basic attributes, financial attributes, and historical behavioral attributes, all derived from data integration results from various internal bank business systems, with a data update frequency of T+1 days. Regarding basic attributes, the system extracts user information such as age, gender, and occupation. Age is a numerical field that has been directly normalized, while gender and occupation are converted into discrete input vectors using a mapping encoding method. For financial attributes, the system extracts key indicators such as total account assets, account opening duration, loan status, wealth management holdings, and fund holdings. The z-score method is used for standardization to eliminate the influence of dimensions, where z = (x - μ) / σ, where μ is the mean and σ is the standard deviation. Regarding historical behavioral attributes, the system aggregates user behavioral statistics for the past 3, 7, 14, and 30 days, covering multiple dimensions such as ad clicks, app activity, logins, purchases, purchase amounts, wealth management purchases, and total wealth management amounts, used to characterize users' medium- to long-term behavioral preferences and financial activity.

[0059] User dynamic behavior features reflect the operational environment and short-term interest state within the current session, characterized by real-time collection and high timeliness. Dynamic features first include contextual attributes; for example, the time of user behavior is divided into multiple hourly segments, including 0–6 AM, 6–9 AM, 9–12 PM, 12–14 PM, 2–17 PM, 5–19 PM, and 7–12 AM, to characterize the temporal periodicity of user behavior. Additionally, it includes geographic location information and network environment (such as 4G, 5G, WiFi, and others) to identify the user's context. Secondly, the system collects the user's page browsing paths and ad click sequences over the past hour. By constructing a dictionary of high-frequency URLs and ad IDs, these behavior sequences are uniformly mapped to fixed-length index sequences, which are then transformed through an embedding layer to form an embedded representation of the behavior sequences. All of the above dynamic behavior data is reported in real-time by the front-end tracking system and written to the Kafka message middleware. By subscribing to the corresponding Topic channel in Kafka, the system can obtain the latest user behavior stream in real-time at minute or shorter intervals, enabling rapid updates and pushes of dynamic features.

[0060] The advertising features mainly include two categories: historical performance features and content features. Historical performance features are statistically consistent with user behavior, calculating the number of impressions, clicks, and corresponding click-through rates over the past 3, 7, 14, and 30 days to assess the recent popularity and attractiveness of the advertisement. Content features include the advertisement title and body text. After preprocessing using the jieba Chinese word segmentation tool, the FinBERT model is used for contextual semantic embedding, transforming it into a unified-dimensional text vector representation. The advertisement also contains structured information, such as the advertisement ID, delivery channel, and delivery time period, all of which undergo discretization encoding and are concatenated with the embedding results to form a complete advertisement feature vector.

[0061] Furthermore, combined Figure 6 This embodiment further illustrates the dynamic behavior feature acquisition, processing, and handling process.

[0062] The system retrieves user behavior data from Kafka real-time streams over the past hour or several time periods, including page browsing paths (URL sequences) and ad click sequences. Each URL in the page browsing behavior is mapped to an integer index based on a predefined vocabulary. The vocabulary retains only the 500 most frequently occurring URLs, while low-frequency URLs are mapped to special tags. Similarly, each ad ID in the ad click behavior is mapped to an integer index. Both types of behavior sequences are processed in a unified format, fixed to a length of 30. If the length is insufficient, zeros are added to the end; if it exceeds the limit, the 30 most recent behaviors are retained. Each behavior element (whether a page URL or an ad ID) is converted into a 16-dimensional vector through a unified embedding layer, forming a uniformly formatted embedding sequence.

[0063] Subsequently, the system calculates a time weight for each embedded vector to reflect the decay effect of user behavior time on short-term interest. The system calculates a time decay factor λ based on the user's daily activity count, using the formula: λ = 1 / (1 + ln(1 + user's daily activity count)). The time weight is obtained through time decay weighting, with the decay formula as follows: Where λ is the time decay factor, and Δt is the browsing duration or click duration in hours. Each embedding vector will be weighted according to this weight to construct a time-sensitive sequence of behavioral representations.

[0064] The processed URL behavior and ad click behavior are embedded and concatenated into a single multimodal behavior sequence, which is then fed into the same GRU model. This GRU model has a temporal recursive structure, enabling it to capture temporal dependencies and higher-order interest transitions in the behavior sequence, and dynamically adjust state updates through a gating mechanism. Finally, the GRU outputs a 128-dimensional sequence representation vector.

[0065] For example: the user's behavior sequence over the past hour is Each action Include:

[0066] Behavioral element: Embedded vector of page URL or ad ID (via embedding layer mapping);

[0067] Timestamp: When the action occurred The interval from the current time is ;

[0068] Embedding vector for each behavior Multiplied by time decay weight ;

[0069] , ;

[0070] The final weighted sequence vector is: ;

[0071] If a user's daily active count is 5, then ;

[0072] Behavior A (10 minutes ago) );

[0073] Behavior B (50 minutes ago) );

[0074] The more active a user is on a given day, the more likely they are to be active. The smaller the interval, the slower the time decay, because the historical behavior of active users needs to be retained for a longer period of time; similarly, the behavior interval... The smaller, then The larger the value, the more important recent behavior is for predicting current interests;

[0075] The weighted sequence Inputting the GRU network yields a 128-dimensional temporal feature vector.

[0076] Meanwhile, the context feature vectors (including time period, geographical location, network environment, etc.) are discretely encoded and then uniformly input into the embedding layer to be converted into a 16-dimensional context embedding representation. For example: the time of occurrence of the behavior is grouped and coded according to time periods, and converted into discrete time period features: (0-6 am is the late night period, coded as 0; 6-9 am is the morning peak, coded as 1; 9-12 am is the morning working time, coded as 2; 12-2 pm is the lunch break, coded as 3; 2-5 pm is the afternoon working time, coded as 4; 5-7 pm is the evening peak, coded as 5; 7-12 am is the nighttime active period, coded as 6); the geographical location is the city where the user is when opening the mobile banking app; the network environment is weighted according to signal strength, with high-quality networks being weighted higher and weak networks being weighted lower, and the influence coefficients corresponding to different network qualities are: {'5G': 1.2, 'WiFi': 1.1, '4G': 1.0, '3G': 0.8, '2G': 0.6, 'Unknown': 1.0}; by concatenating the above context features and inputting them into the embedding layer, the embedding layer converts them into a 16-dimensional vector. Finally, the 16-dimensional context vector is concatenated with the aforementioned 128-dimensional GRU output vector in the feature dimension, that is, a dynamic behavior feature vector with a length of 144 dimensions is formed through vector concatenation operation.

[0077] S200: The user offline features and the set of advertising features are combined one by one and input into the pre-trained first model to generate the first set of prediction scores.

[0078] In this embodiment, the user's offline characteristics mainly include the user's basic attributes, financial attributes, and historical behavioral statistics. Basic attributes include static information such as the user's age, gender, and occupation; financial attributes include structured indicators from the bank's core system and wealth management system, such as the user's current total assets, account opening duration, loan status, wealth management holdings, and number of fund holdings; historical behavioral statistics include the user's ad clicks, login activity, consumption frequency, consumption amount, and number and amount of wealth management purchases in the past 3, 7, 14, and 30 days. All of the above offline characteristics are structured data with a T+1 time update granularity, and are typically synchronized daily.

[0079] The ad feature set includes statistical information such as historical exposure counts, click counts, and click-through rates for candidate ads across different time windows. It also includes metadata such as ad content type coding, risk tags, and delivery channels. These ad features are extracted and standardized from the ad delivery system through an offline processing workflow.

[0080] During feature input, the system pairs each user's offline features with the features of each candidate ad to be delivered, constructing a joint feature vector. This joint feature vector is then fed into the pre-trained first prediction model for batch inference. This first model is a structured gradient boosting tree model (such as XGBoost), capable of processing sparse, high-dimensional structured data and effectively modeling the nonlinear relationship between ad features and long-term user preferences.

[0081] S300. The user dynamic features and the set of advertising features are combined one by one and input into the pre-trained second model to generate a second set of prediction scores.

[0082] In this embodiment, the user dynamic features consist of two parts: contextual features and sequential behavior features. Contextual features include the time period in which the user's behavior occurred, their geographical location at that time, and the network environment status, ultimately mapped to a 16-dimensional vector. Sequential behavior features include the page browsing path and ad click history within the past hour, fed into a shared GRU network for modeling, ultimately outputting a unified 128-dimensional behavior vector representing the user's short-term preferences. The contextual embedding vector (16-dimensional) and the GRU behavior encoding (128-dimensional) are concatenated to obtain a final dynamic feature vector of length 144 dimensions.

[0083] The processing method for the ad feature set is consistent with that in step S200, that is, the content features, historical exposure and click data, and other information of each candidate ad constitute an input pair combined with each user's dynamic features. These dynamic features are then combined with the ad features one by one and input into a pre-trained second model. This second model is a deep behavior prediction model (such as DeepFM) with embedding modeling capabilities, where the FM part is used to capture the cross-relationships between features, and the DNN part is used to capture non-linear combinations.

[0084] The model output is the second predicted score, representing the user's short-term interest in the advertisement under the current context and behavioral drive. Repeating the above process for all candidate advertisements generates a set of second predicted scores, which serves as input for subsequent weighted fusion.

[0085] S400. The first predicted score set and the second predicted score set are weighted and fused to obtain a fused predicted score set.

[0086] In this embodiment, the first predicted score set represents the user's long-term static preference tendency, derived from the output of the XGBoost model with the combination of user offline features and advertising features as input; the second predicted score set reflects the user's short-term behavioral interests, derived from the output of the DeepFM model with the combination of user dynamic features and advertising features as input. To more comprehensively characterize the user's overall interest in advertising, a weighted fusion strategy is introduced to integrate the prediction results of the two models.

[0087] Before fusion, the system first identifies the user type of the current users, classifying them into new users, active users, or low-frequency users according to preset rules. New users are those who have been registered for less than 30 days; active users are those who have visited the platform more than 3 times per day on average in the past 7 days; and low-frequency users are other users who do not meet the above criteria. User type identification is a prerequisite for the fusion strategy, used to dynamically determine the fusion weights of the scores of the two models.

[0088] The system pre-sets a set of weighted coefficient matrices, with different fusion strategies corresponding to different user types. Taking active users as an example, their behavioral data is abundant, and their short-term interests are more pronounced; therefore, the system assigns weighted coefficients of 0.4 for the first model score and 0.6 for the second model score. For new users, due to less behavioral data and more stable and reliable static features, the weighted coefficients are adjusted to 0.7 for the first model score and 0.3 for the second model score. All of the above weight values ​​can be optimized offline through A / B testing using historical data. In one embodiment, the list of user type weighted coefficients is as follows:

[0089]

[0090] For new users (α=0.7, β=0.3): Rely on static profiles (XGBoost dominant). For active users (α=0.4, β=0.6): Focus on real-time behavior (DeepFM dominant). For example, if a customer has an XGBoost score of 0.8 and a DeepFM score of 0.5, and this customer is an active user, then their combined score is 0.4x0.8 + 0.6x0.5 = 0.62.

[0091] S500. Based on the compliance adjustment coefficient and the fusion prediction score set, the final advertising score set is obtained.

[0092] In this embodiment, to achieve compliance control over advertising content delivery, the system constructs a compliance control matrix based on the correspondence between user risk levels and advertising risk types, and performs non-linear corrections on the fused prediction score set accordingly to obtain the final advertising score set. This step aims to organically integrate the personalized recommendation mechanism with bank risk control requirements, preventing high-risk advertisements from being pushed to unsuitable users and ensuring that the system improves click-through rates without triggering regulatory compliance risks.

[0093] First, the system sets multiple risk levels on the user side, including at least three categories: conservative, moderate, and aggressive. These risk levels are derived from a comprehensive range of factors, including the user's asset structure, age, past financial product preferences, and questionnaire assessment results. Simultaneously, for each candidate advertisement, a risk label and type are assigned, categorized into three main types: product, activity, and equity, further subdivided into low, medium, and high risk levels, forming a complete advertising risk classification system.

[0094] After the system obtains the current user's fusion prediction score set, it first determines the user's risk level; then, for each candidate advertisement, it identifies its risk type. The system then queries the compliance control matrix using "user risk level × advertisement risk level" as the key to obtain a compliance adjustment coefficient r. This compliance adjustment coefficient is a floating-point number between 0 and 1, used to suppress mismatched placements and ensure the effectiveness of risk control rules. For example, if a conservative user encounters a high-risk financial advertisement, the adjustment coefficient r might be 0.3, while a moderate user encountering a medium-risk activity advertisement might have r at 0.7.

[0095] In one embodiment, the compliance adjustment coefficients are listed below:

[0096] User Risks / Advertising Types Equity e-commerce Events Product Category - Low Product Category - Medium Product Category - High R1 Conservative 1.0 0.9 0.8 0.7 0.5 0.3 R2 Stable 1.0 1.0 0.9 0.8 0.7 0.5 R3 Balanced Type 1.0 1.0 1.0 0.9 0.8 0.7 R4 Growth Type 1.0 1.0 1.0 1.0 0.9 0.8 R5 Progressive 1.0 1.0 1.0 1.0 1.0 1.0

[0097] A compliance adjustment coefficient will be assigned based on different risk levels of customers and advertising types; the invention employs a two-factor integration mechanism: Where α: business factor (0.7-1.2), controlling the contribution strength of the fusion score; β: compliance factor (0.8-1.5), controlling the sensitivity of compliance impact. Design advantages: when α>1, it amplifies the prediction weight of high-value users; when β>1, it enhances the compliance impact of high-risk advertisements.

[0098] For example, when α=1.2, the predicted score for high-value users (such as VIPs) is amplified by 1.2 times. When β=1.5, the score for high-risk advertisements (such as R1 users recommending high-risk financial products) becomes 1.5 times the original value: querying the compliance coefficient matrix yields a compliance coefficient of 0.3, and 0.3 to the power of 1.5 is approximately equal to 0.164, indicating increased compliance suppression, and the score drops from 0.3 to 0.164.

[0099] To prevent overexposure (preventing high-scoring ads from being overexposed and maintaining ad diversity) and protect long-tail ads (ensuring basic exposure for low-scoring ads and supporting the cold start of new products), the following business protection constraints are added: If the calculated final delivery score is >0.95, then the final delivery score = min(1.0, final delivery score * 0.98); if the calculated final delivery score is <0.3, then the final delivery score = max(0.15, final delivery score * 1.05).

[0100] When an ad score is too high (>0.95), it indicates that the system excessively favors that ad. A fine-tuning adjustment is made by multiplying by 0.98 (attenuating by 2%) to reduce its weight, and min(1.0, ...) ensures the score does not exceed the upper limit of 1.0. From a business value perspective: 1) Avoid a single ad occupying >30% of exposure resources; 2) Increase the exposure opportunities of long-tail ads by 5-8%; 3) Reduce user ad fatigue rate (actually reduced by 12%). Scenarios are as follows:

[0101] Scene Raw score Adjusted score Effect Popular financial advertisements 0.97 min(1.0, 0.97×0.98=0.9506) Reduced weight to 0.9506 Popular credit card ads 0.99 min(1.0, 0.99×0.98=0.9702) Reduced weight to 0.9702 Threshold advertising 0.96 min(1.0, 0.96×0.98=0.9408) Reduced weight to 0.9408 Ads that have reached their limit 1.00 min(1.0, 1.00×0.98=0.98) Reduced weighting to 0.98

[0102] When an ad score is too low (<0.3), it indicates that the system is excessively ignoring the ad. Multiply the score by 1.05 (5% increase) for basic weighting, and set a minimum protection line using max(0.15, ...). From a business value perspective: 1) Ensures new ads receive at least 0.5% basic exposure; 2) Increases long-tail ad click-through rates by 15-20%; 3) Shortens the cold start cycle for new products by 3-5 days. Scenarios are as follows:

[0103] Scene Raw score Adjusted score Effect Newly launched insurance products 0.28 max(0.15, 0.28×1.05=0.294) Increased to 0.294 Niche Forex Services 0.25 max(0.15, 0.25×1.05=0.2625) Increased to 0.2625 Low-yield government bonds 0.10 max(0.15, 0.10×1.05=0.105) Increase to 0.15 (to ensure a minimum). Discounts for less popular merchants 0.18 max(0.15, 0.18×1.05=0.189) Increased to 0.189

[0104] S600: Select at least one advertisement with the highest score from the final advertisement score set and place it in the mobile banking banner position.

[0105] In step S600, after receiving the final ad score set adjusted for compliance factors, the system enters the ad placement decision stage. The final ad score set for each user is in structured data format, containing a unique identifier for each candidate ad and its corresponding final score. This score incorporates model predictions based on both static and dynamic features, and has undergone non-linear correction using compliance adjustment coefficients. The system first sorts the ad scores in the set in descending order, resulting in an ad priority queue arranged from highest to lowest user suitability. Then, based on the actual capacity configuration of the banner display slots, the system selects the top-scoring ads as the targets for placement. If the banner area uses a single ad display mode, the system directly selects the ad with the highest score; if a carousel mode is supported, the top-scoring ads are sequentially added. The finally selected ads are displayed on the mobile banking app's homepage banner area via the rendering engine. Simultaneously, the system records information such as the user identifier, ad identifier, corresponding score, and placement time for subsequent click-through rate evaluation, model iteration training, and performance review. This step completes the technical loop from feature acquisition, model prediction, score fusion, compliance correction to final ad selection and display, enabling precise ad delivery and dynamic control for mobile banking users.

[0106] Furthermore, in this embodiment, the first model is a structured gradient boosting prediction model, used to receive a joint input vector composed of user static features and advertising features, and output a first prediction score. The first model is a structured gradient boosting prediction model, specifically employing ensemble learning algorithms such as XGBoost and LightGBM, which possess feature importance assessment and efficient training capabilities. This model processes structured static input features. The input vector is composed of a combination of the user's offline static features and advertising features. The static features include basic user attributes, financial attributes, and historical behavioral statistics, while the advertising features include advertising content, advertising type, and historical exposure and click performance. The model learns by training on the above feature space and outputs a first prediction score for each advertisement, reflecting the user's long-term advertising preference trend without a clear contextual behavioral background.

[0107] The second model is an embedded behavioral modeling-based deep prediction model, which receives a joint input vector composed of user dynamic behavior features and advertising features, and outputs a second prediction score. This second model is preferably a DeepFM model based on a deep fusion structure of FM and DNN. The input to this model is a joint vector of user dynamic behavior features and advertising features. User dynamic behavior features include contextual features such as geographical location, network environment, and access time period, and behavioral sequence features such as the user's recent page browsing path and advertising click sequence. After encoding, time decay weighting, and GRU encoding, these features are formed into an embedded feature representation. Advertising features also participate in cross-modeling and deep learning through embedding, ultimately outputting a second prediction score to characterize the user's immediate interest in different advertisements in the current behavioral context.

[0108] In summary, the predicted scores generated by the first model focus on reflecting users' long-term static preferences, exhibiting advantages in stability and interpretability. In contrast, the predicted scores generated by the second model emphasize modeling users' short-term behavioral interests and their ability to respond instantly. These two models, approaching the issue from different feature perspectives, constitute a complementary dual-path advertising scoring system, providing high-quality foundational scores for subsequent fusion predictions.

[0109] Figure 2 This is a schematic diagram illustrating a fusion prediction score calculation method provided in an embodiment of this disclosure. Now, regarding... Figure 2 Let's elaborate on that.

[0110] S410. Classify the current users into at least three categories: new users, active users, and low-frequency users.

[0111] S410. Classify current users into at least three categories: new users, active users, and low-frequency users. Specifically, the system models user groups based on metrics such as registration duration, recent active days, and access frequency. New users are those registered for less than a preset threshold (e.g., 7 days) and with limited historical behavioral data; active users are those with high access frequency and numerous actions in the recent period (e.g., within the last 7 days); and low-frequency users are those with low access frequency and sparse interaction behavior in the recent period. The results of this user classification will serve as the basis for subsequent weighted fusion strategies, used to achieve differentiated fusion of prediction scores among different user groups, thereby improving the personalization and response accuracy of recommendation results.

[0112] S420. Set different weighting coefficients for the first predicted score set and the second predicted score set according to different user types.

[0113] Specifically, after segmenting users, the system sets differentiated weighting strategies based on the varying degrees of dependence on static and dynamic features for each user group. For example, for new users, due to their limited historical behavior and scarce dynamic features, the system prioritizes the first predicted score generated by static features, thus assigning a higher weight to the first score (e.g., 0.8), while the weight of the second score is correspondingly reduced (e.g., 0.2). For active users, whose behavior is richer and whose dynamic features better represent their current interests, the system appropriately increases the weight of the second predicted score (e.g., the first score weight is 0.4, and the second score weight is 0.6). For low-frequency users, an intermediate weight allocation is set based on a combination of their historical static attributes and recent limited behavior (e.g., the first score weight is 0.6, and the second score weight is 0.4). The weighting coefficients can be periodically updated or dynamically adjusted based on the A / B test results of real user click behavior to ensure the adaptability and robustness of the fusion strategy.

[0114] S430. The first predicted score set and the second predicted score set are weighted and fused according to the weighting coefficients to obtain the fused predicted score set.

[0115] The system performs a weighted fusion of the first predicted score set and the second predicted score set based on the weighting coefficients to obtain the fused predicted score set. Specifically, for each candidate advertisement, the first predicted score generated by the first model is extracted. and the second prediction score generated by the second model The corresponding weighting coefficient is determined based on the current user type (e.g., new user, active user, low-frequency user). and ,Require and The sum is 1. Then, a linear weighted fusion operation is performed, and the fusion score, score_fusion, is calculated as score_fusion = × + × . It mainly reflects the user's long-term static preferences. This reflects users' short-term behavioral interests. Through the weighted fusion method described above, while fully preserving the output information of both prediction models, differentiated fusion weights are set based on the behavioral characteristics of different user groups. After the fusion scores of all candidate ads are calculated using the above method, they constitute the fusion prediction score set corresponding to the current user, serving as the input basis for subsequent compliance adjustments and final ad selection. This strategy can more accurately match the evolution trend of users' advertising interests, improving the real-time performance and relevance of personalized recommendations.

[0116] Figure 3This is a schematic diagram illustrating a final advertising score calculation method provided in an embodiment of this disclosure. Now, in conjunction with... Figure 3 The specific embodiments of this application are further described below.

[0117] S510. Based on the current user's risk level and the risk type of the candidate advertisement, query the preset compliance control matrix to obtain the compliance adjustment coefficient corresponding to each fusion prediction score.

[0118] The system, based on the current user's risk level and the risk type of candidate advertisements, queries a pre-defined compliance control matrix to obtain the compliance adjustment coefficient corresponding to each fusion prediction score. Specifically, the system first categorizes user risk levels into conservative, stable, and aggressive types based on factors such as asset size, investment preferences, and historical trading behavior. Advertisement content is then categorized into various risk types, such as product-based, activity-based, and equity-based, based on product characteristics, security sensitivity, and user suitability requirements. The system combines the user's current risk level with the risk type of candidate advertisements to form a two-dimensional coordinate index, used to locate and extract the compliance adjustment coefficient for each advertisement within the compliance control matrix. This adjustment coefficient is a floating-point number, typically limited to 0 to 1, reflecting the strength of compliance recommendation for a particular user type when faced with a particular type of advertisement content. Through this matrix query operation, the system can achieve risk adaptation control at the user-advertisement level, ensuring that subsequent advertisement scoring has differentiated, interpretable, and compliance-oriented characteristics, providing basic parameters for the next step of non-linear adjustment calculations.

[0119] S520. For each score value in the fusion prediction score set, apply the compliance adjustment coefficient to perform nonlinear calculation to generate the final advertising score set.

[0120] The system applies the compliance adjustment coefficient to each score value in the fusion prediction score set for nonlinear calculation, generating the final ad score set. This process no longer uses simple linear weighting or proportional scaling, but introduces a nonlinear mapping function to reprocess the score values, giving the impact of risk control greater discriminative and regulatory capabilities. Under the nonlinear processing mechanism, when the compliance adjustment coefficient for a user and ad combination is low, it indicates a high compliance risk, and the original fusion score will be significantly reduced to decrease the ad's display probability. Conversely, a high compliance adjustment coefficient indicates a low ad risk, and the system retains a higher degree of the original fusion score, thus maintaining ad effectiveness. In this nonlinear processing, the system can also introduce parameters such as risk suppression factors and classification adjustment strategies to achieve refined compliance scoring for different user levels and ad types, thereby improving delivery efficiency and user experience while ensuring compliance.

[0121] Figure 4This is a schematic diagram illustrating a compliance control matrix construction method provided in an embodiment of this disclosure. Now, in conjunction with... Figure 4 The specific embodiments of this application are further described below.

[0122] S1. Preset multiple user risk levels, including at least conservative, moderate and aggressive.

[0123] The system comprehensively assesses users' risk tolerance based on factors such as their account asset structure, investment preferences, historical purchase records, age group, and regulatory requirements within mobile banking, and classifies them into risk levels accordingly. Conservative users primarily focus on principal security, tend to choose low-risk products, and have low tolerance for risky advertisements; moderate users pursue returns while possessing a certain risk tolerance, showing a high tolerance for low- to medium-risk advertisements; aggressive users prefer high-yield, high-risk products, have a wider range of advertising acceptance, and are willing to participate in benefit activities and try new services. This risk level setting provides the foundation for subsequently building a compliance-adaptive relationship between "users" and "advertisements." The system can achieve periodic updates and accurate identification of user risk levels through offline profiling modeling or by combining external compliance data.

[0124] S2. Configure multiple risk tags for candidate ads, including at least product category, event category, and benefit category, and their corresponding risk levels.

[0125] Before ad delivery, the system conducts content review and risk assessment on all candidate ads, categorizing them based on the financial product attributes, promotional methods, and customer impact. Product ads typically contain information on financial products such as wealth management, funds, and insurance, carrying certain investment risks. The system further refines the risk level based on product risk ratings (e.g., R1-R5). Activity ads include points redemption, limited-time offers, and marketing promotions; while lower risk, they may involve frequent interactions or user fund transfers. Benefit ads involve VIP privileges and upgrades; while relatively low risk, they involve identity verification and restrictions on benefit usage. Each ad category is assigned a corresponding risk level tag based on its potential impact on user decision-making and fund security. This tag is then used for subsequent matching analysis with the user's risk level to achieve personalized ad recommendations with controllable risk.

[0126] S3. Combine user risk level and advertising risk level in pairs to construct a matrix coordinate space, and configure a corresponding compliance adjustment coefficient for each combination. The compliance adjustment coefficient is a floating-point number between 0 and 1.

[0127] Specifically, the system uses user risk level as the horizontal axis and ad risk level as the vertical axis to form a two-dimensional matrix space. Each coordinate point represents a user-ad risk matching combination. Within this coordinate space, the system presets a compliance adjustment coefficient for each combination to adjust the final weight of the fusion prediction score. The closer the adjustment coefficient is to 1, the higher the user's risk matching with that type of ad, and the stronger the recommendation intention; conversely, the closer the coefficient is to 0, the greater the risk mismatch, and the system will significantly reduce its display probability. This compliance adjustment coefficient acts as a multiplicative factor on the fusion prediction score during model execution, achieving dynamic compliance calibration of the ad recommendation results, thereby improving the advertising system's ability to coordinate compliance and personalized recommendations.

[0128] Figure 5 This is a schematic diagram of a mobile banking banner ad intelligent delivery system based on a fusion model, provided as an embodiment of this disclosure. Now, in conjunction with... Figure 5 The specific embodiments of this application are further described below.

[0129] The feature acquisition module is used to acquire user offline features, user dynamic features, and advertising feature sets.

[0130] User offline characteristics mainly include basic attributes, financial attributes, and historical behavioral attributes, all derived from data integration results from various internal bank business systems, with a data update frequency of T+1 days. Regarding basic attributes, the system extracts user information such as age, gender, and occupation. Age is a numerical field that has been directly normalized, while gender and occupation are converted into discrete input vectors using a mapping encoding method. For financial attributes, the system extracts key indicators such as total account assets, account opening duration, loan status, wealth management holdings, and fund holdings. The z-score method is used for standardization to eliminate the influence of dimensions, where z = (x - μ) / σ, where μ is the mean and σ is the standard deviation. Regarding historical behavioral attributes, the system aggregates user behavioral statistics for the past 3, 7, 14, and 30 days, covering multiple dimensions such as ad clicks, app activity, logins, purchases, purchase amounts, wealth management purchases, and total wealth management amounts, used to characterize users' medium- to long-term behavioral preferences and financial activity.

[0131] User dynamic behavior features reflect the operational environment and short-term interest state within the current session, characterized by real-time collection and high timeliness. Dynamic features first include contextual attributes; for example, the time of user behavior is divided into multiple hourly segments, including 0–6 AM, 6–9 AM, 9–12 PM, 12–14 PM, 2–17 PM, 5–19 PM, and 7–12 AM, to characterize the temporal periodicity of user behavior. Additionally, it includes geographic location information and network environment (such as 4G, 5G, WiFi, and others) to identify the user's context. Secondly, the system collects the user's page browsing paths and ad click sequences over the past hour. By constructing a dictionary of high-frequency URLs and ad IDs, these behavior sequences are uniformly mapped to fixed-length index sequences, which are then transformed through an embedding layer to form an embedded representation of the behavior sequences. All of the above dynamic behavior data is reported in real-time by the front-end tracking system and written to the Kafka message middleware. By subscribing to the corresponding Topic channel in Kafka, the system can obtain the latest user behavior stream in real-time at minute or shorter intervals, enabling rapid updates and pushes of dynamic features.

[0132] The advertising features mainly include two categories: historical performance features and content features. Historical performance features are statistically consistent with user behavior, calculating the number of impressions, clicks, and corresponding click-through rates over the past 3, 7, 14, and 30 days to assess the recent popularity and attractiveness of the advertisement. Content features include the advertisement title and body text. After preprocessing using the jieba Chinese word segmentation tool, the FinBERT model is used for contextual semantic embedding, transforming it into a unified-dimensional text vector representation. The advertisement also contains structured information, such as the advertisement ID, delivery channel, and delivery time period, all of which undergo discretization encoding and are concatenated with the embedding results to form a complete advertisement feature vector.

[0133] The first scoring module is used to combine the user's offline features and the set of advertising features one by one and input them into the pre-trained first model to generate a first set of predicted scores.

[0134] In this embodiment, the user's offline characteristics mainly include the user's basic attributes, financial attributes, and historical behavioral statistics. Basic attributes include static information such as the user's age, gender, and occupation; financial attributes include structured indicators from the bank's core system and wealth management system, such as the user's current total assets, account opening duration, loan status, wealth management holdings, and number of fund holdings; historical behavioral statistics include the user's ad clicks, login activity, consumption frequency, consumption amount, and number and amount of wealth management purchases in the past 3, 7, 14, and 30 days. All of the above offline characteristics are structured data with a T+1 time update granularity, and are typically synchronized daily.

[0135] The ad feature set includes statistical information such as historical exposure counts, click counts, and click-through rates for candidate ads across different time windows. It also includes metadata such as ad content type coding, risk tags, and delivery channels. These ad features are extracted and standardized from the ad delivery system through an offline processing workflow.

[0136] During feature input, the system pairs each user's offline features with the features of each candidate ad to be delivered, constructing a joint feature vector. This joint feature vector is then fed into the pre-trained first prediction model for batch inference. This first model is a structured gradient boosting tree model (such as XGBoost), capable of processing sparse, high-dimensional structured data and effectively modeling the nonlinear relationship between ad features and long-term user preferences.

[0137] The second scoring module is used to combine the user dynamic features and the set of advertising features one by one and input them into the pre-trained second model to generate a second set of predicted scores.

[0138] In this embodiment, the user dynamic features consist of two parts: contextual features and sequential behavior features. Contextual features include the time period in which the user's behavior occurred, their geographical location at that time, and the network environment status, ultimately mapped to a 16-dimensional vector. Sequential behavior features include the page browsing path and ad click history within the past hour, fed into a shared GRU network for modeling, ultimately outputting a unified 128-dimensional behavior vector representing the user's short-term preferences. The contextual embedding vector (16-dimensional) and the GRU behavior encoding (128-dimensional) are concatenated to obtain a final dynamic feature vector of length 144 dimensions.

[0139] The processing method for the ad feature set is consistent with that in step S200, that is, the content features, historical exposure and click data, and other information of each candidate ad constitute an input pair combined with each user's dynamic features. These dynamic features are then combined with the ad features one by one and input into a pre-trained second model. This second model is a deep behavior prediction model (such as DeepFM) with embedding modeling capabilities, where the FM part is used to capture the cross-relationships between features, and the DNN part is used to capture non-linear combinations.

[0140] The model output is the second predicted score, representing the user's short-term interest in the advertisement under the current context and behavioral drive. Repeating the above process for all candidate advertisements generates a set of second predicted scores, which serves as input for subsequent weighted fusion.

[0141] The fusion score module is used to perform weighted fusion of the first predicted score set and the second predicted score set to obtain a fused predicted score set.

[0142] In this embodiment, the first predicted score set represents the user's long-term static preference tendency, derived from the output of the XGBoost model with the combination of user offline features and advertising features as input; the second predicted score set reflects the user's short-term behavioral interests, derived from the output of the DeepFM model with the combination of user dynamic features and advertising features as input. To more comprehensively characterize the user's overall interest in advertising, a weighted fusion strategy is introduced to integrate the prediction results of the two models.

[0143] Before fusion, the system first identifies the user type of the current users, classifying them into new users, active users, or low-frequency users according to preset rules. New users are those registered less than 30 days ago; active users are those with an average of more than 3 visits per day in the past 7 days; and low-frequency users are other users who do not meet the above criteria. User type identification is a prerequisite for the fusion strategy, used to dynamically determine the fusion weights of the two model scores. The system presets a set of weighted coefficient matrices, with different fusion strategies corresponding to different user types.

[0144] The final scoring module is used to obtain the final advertising score set based on the compliance adjustment coefficient and the fusion prediction score set.

[0145] In this embodiment, to achieve compliance control over advertising content delivery, the system constructs a compliance control matrix based on the correspondence between user risk levels and advertising risk types, and performs non-linear corrections on the fused prediction score set accordingly to obtain the final advertising score set. This step aims to organically integrate the personalized recommendation mechanism with bank risk control requirements, preventing high-risk advertisements from being pushed to unsuitable users and ensuring that the system improves click-through rates without triggering regulatory compliance risks.

[0146] First, the system sets multiple risk levels on the user side, including at least three categories: conservative, moderate, and aggressive. These risk levels are derived from a comprehensive range of factors, including the user's asset structure, age, past financial product preferences, and questionnaire assessment results. Simultaneously, for each candidate advertisement, a risk label and type are assigned, categorized into three main types: product, activity, and equity, further subdivided into low, medium, and high risk levels, forming a complete advertising risk classification system.

[0147] After the system obtains the current user's fusion prediction score set, it first determines the user's risk level; then, for each candidate advertisement, it identifies its risk type. The system then queries the compliance control matrix using "user risk level × advertisement risk level" as the key to obtain a compliance adjustment coefficient r. This compliance adjustment coefficient is a floating-point number between 0 and 1, used to suppress mismatched placements and ensure the effectiveness of risk control rules. For example, if a conservative user encounters a high-risk financial advertisement, the adjustment coefficient r might be 0.3, while a moderate user encountering a medium-risk activity advertisement might have r at 0.7.

[0148] A compliance adjustment coefficient will be assigned based on different risk levels of customers and advertising types; the invention employs a two-factor integration mechanism: Where α: business factor (0.7-1.2), controlling the contribution strength of the fusion score; β: compliance factor (0.8-1.5), controlling the sensitivity of compliance impact. Design advantages: when α>1, it amplifies the prediction weight of high-value users; when β>1, it enhances the compliance impact of high-risk advertisements.

[0149] For example, when α=1.2, the predicted score for high-value users (such as VIPs) is amplified by 1.2 times. When β=1.5, the score for high-risk advertisements (such as R1 users recommending high-risk financial products) becomes 1.5 times the original value: querying the compliance coefficient matrix yields a compliance coefficient of 0.3, and 0.3 to the power of 1.5 is approximately equal to 0.164, indicating increased compliance suppression, and the score drops from 0.3 to 0.164.

[0150] The ad placement module is used to select at least one ad with the highest score from the final ad score set and place it in the mobile banking banner position.

[0151] After receiving the final ad score set adjusted for compliance factors, the system enters the ad placement decision-making stage. The final ad score set for each user is in structured data format, containing a unique identifier for each candidate ad and its corresponding final score. This score incorporates model predictions based on both static and dynamic features, and has undergone non-linear correction using a compliance adjustment coefficient. The system first sorts the ad scores in the set in descending order, resulting in an ad priority queue arranged from highest to lowest user suitability. Then, based on the actual capacity of the banner display slots, the system selects the top-scoring ads as targets for placement. If the banner area uses a single-ad display mode, the system directly selects the ad with the highest score; if a carousel mode is supported, the top-scoring ads are sequentially added. The finally selected ads are displayed on the mobile banking app's homepage banner area via the rendering engine. Simultaneously, the system records information such as the user identifier, ad identifier, corresponding score, and placement time for subsequent click-through rate evaluation, model iteration training, and performance review. This step completes the technical loop from feature acquisition, model prediction, score fusion, compliance correction to final ad selection and display, enabling precise ad delivery and dynamic control for mobile banking users.

[0152] Furthermore, in one embodiment, the system is deployed on a server cluster with heterogeneous model processing capabilities. The server cluster includes general-purpose CPU nodes for executing the first model prediction and GPU nodes for executing the second model. Specifically, the first model is a structured gradient enhancement model with relatively low computational complexity, suitable for efficient inference in batch processing on CPU nodes; the second model is an embedded behavioral modeling deep neural network containing embedding layers, GRU sequence layers, and multi-layer fully connected networks, with high computational load and strong parallelism, making it more suitable for deployment on GPU nodes that support tensor computation for real-time prediction. By heterogeneously scheduling and allocating model inference tasks in the server cluster, not only is the overall response efficiency and throughput of the system improved, but the stability and scalability of the system are also ensured when facing large-scale user behavior data and diverse advertising requests. This heterogeneous deployment scheme significantly reduces the risk of computing resource conflicts and latency, providing hardware support for the system's real-time performance and high-concurrency processing capabilities.

[0153] It should be noted that the acquisition, storage, use, and processing of data in this disclosed technical solution comply with the relevant provisions of national laws and regulations. The various types of data, such as personal identity data, operational data, and behavioral data related to individuals, customers, and groups, obtained in this disclosed embodiment have all been authorized.

[0154] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute a configuration software-based software licensing implementation method.

[0155] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0156] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the configuration software-based software licensing implementation methods provided by the above methods.

[0157] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0159] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method for intelligent delivery of mobile banking banner ads based on a fusion model, characterized in that, include: Obtain a set of user offline characteristics, user dynamic characteristics, and advertising characteristics; The user offline features and the set of advertising features are combined one by one and input into the pre-trained first model to generate the first set of prediction scores. The user dynamic features and the set of advertising features are combined one by one and input into the pre-trained second model to generate a second set of prediction scores. The first predicted score set and the second predicted score set are weighted and fused to obtain a fused predicted score set. The final advertising score set is obtained by combining the compliance adjustment coefficient with the fusion prediction score set. Select at least one advertisement with the highest score from the final ad score set and place it in the mobile banking banner position; The first model is a structured gradient enhancement prediction model, which receives a joint input vector consisting of the user's static features and the advertising features, and outputs a first prediction score. The second model is an embedded behavior modeling deep prediction model, which is used to receive a joint input vector composed of the user's dynamic behavior features and advertising features, and output a second prediction score. The first prediction score and the second prediction score reflect the user's long-term static preferences and short-term behavioral interests, respectively.

2. The intelligent delivery method for mobile banking banner ads based on a fusion model according to claim 1, characterized in that, The step of weightedly fusing the first predicted score set and the second predicted score set to obtain a fused predicted score set further includes: The current users are categorized into at least three types: new users, active users, and low-frequency users. Different weighting coefficients are set for the first predicted score set and the second predicted score set according to different user types; The first predicted score set and the second predicted score set are weighted and fused according to the weighting coefficients to obtain the fused predicted score set.

3. The intelligent delivery method for mobile banking banner ads based on a fusion model according to claim 1, characterized in that, The step of obtaining the final advertising score set based on the compliance adjustment coefficient and the fusion prediction score set also includes: Based on the current user's risk level and the risk type of the candidate advertisement, query the preset compliance control matrix to obtain the compliance adjustment coefficient corresponding to each fusion prediction score; For each score value in the fusion prediction score set, the compliance adjustment coefficient is applied to perform a nonlinear calculation to generate the final advertising score set.

4. The intelligent delivery method for mobile banking banner ads based on a fusion model according to claim 1, characterized in that, The pre-trained first model and the pre-trained second model are configured with update strategies, including: The first model performs retraining on the full dataset daily and replaces the model on a daily schedule. The second model uses an incremental update mechanism, acquiring new user behavior data every hour to fine-tune the model; During the update process of the second model, the cross feature parameters in the FM layer are iteratively updated using the stochastic gradient descent method, the embedding layer and intermediate hidden layer of the DNN layer are frozen, and only the last two layers are fine-tuned.

5. The intelligent delivery method for mobile banking banner ads based on a fusion model according to claim 3, characterized in that, The construction of the compliance control matrix includes: Multiple user risk levels are preset, including at least conservative, moderate, and aggressive risk levels; Configure multiple risk tags for candidate ads, including at least product category, event category, and benefit category, and their corresponding risk levels; The user risk level and the advertising risk level are combined in pairs to construct a matrix coordinate space, and a corresponding compliance adjustment coefficient is configured for each combination. The compliance adjustment coefficient is a floating-point number between 0 and 1.

6. A mobile banking banner ad intelligent delivery system based on a fusion model, characterized in that, include: The feature acquisition module is used to acquire user offline features, user dynamic features, and advertising feature sets; The first scoring module is used to combine the user's offline features with the set of advertising features one by one and input them into the pre-trained first model to generate a first set of predicted scores. The second scoring module is used to combine the user dynamic features and the set of advertising features one by one and input them into the pre-trained second model to generate a second set of predicted scores. The fusion score module is used to perform weighted fusion of the first predicted score set and the second predicted score set to obtain a fused predicted score set; The final scoring module is used to obtain the final advertising score set based on the compliance adjustment coefficient and the fusion prediction score set; The ad delivery module is used to select at least one ad with the highest score from the final ad score set and place it in the mobile banking banner position; The first model is a structured gradient enhancement prediction model, which is used to receive a joint input vector composed of the user's static features and the advertising features, and output a first prediction score. The second model is an embedded behavior modeling deep prediction model, which is used to receive a joint input vector composed of the user's dynamic behavior features and advertising features, and output a second prediction score. The first prediction score and the second prediction score reflect the user's long-term static preferences and short-term behavioral interests, respectively.

7. The intelligent mobile banking banner ad delivery system based on a fusion model according to claim 6, characterized in that, The system is deployed on a server cluster with heterogeneous model processing capabilities. The server cluster includes general-purpose CPU nodes for performing the first model prediction and GPU nodes for performing the second model.

8. An electronic device, comprising: A processor, and a memory communicatively connected to the processor; characterized in that: The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory, in accordance with the steps of the method according to any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-5.

Images