Multi-type advertisement recommendation method and system based on user portrait

Abstract

This invention provides a method and system for recommending multiple types of advertisements based on user profiles, relating to the field of advertising delivery technology. The method includes: analyzing the temporal dependencies between behaviors based on historical user behavior data across multiple interaction channels, under a privacy protection mechanism; performing probabilistic behavior analysis based on these dependencies; generating an advertisement contact sequence arranged according to a pre-defined narrative logic based on a probabilistic user profile and current context information; matching corresponding advertisement content according to the advertisement type and narrative requirements of each node in the sequence; and delivering and pushing advertisements to user terminals according to the temporal relationship of the sequence. This solves the technical problem of existing technologies where one-sided user behavior analysis leads to low advertisement recommendation accuracy, affecting advertising delivery effectiveness and user experience. It achieves accurate recommendation of multiple types of advertisements, improving advertising delivery effectiveness and user experience.

Description

Technical Field

[0001] This invention relates to the field of advertising delivery technology, specifically to a method and system for recommending multiple types of advertisements based on user profiles. Background Technology

[0002] Current ad recommendations primarily rely on content-based recommendation or collaborative filtering methods, combining users' historical clicks, browsing, or purchase behavior to push ads that reflect their potential interests. These methods typically consider only a single ad type or independent behavioral characteristics, ignoring the temporal relationships of user behavior and the characteristics of multi-channel interaction. This leads to recommendations that are prone to duplication or do not match actual user preferences. Furthermore, current ad recommendations often employ static rules or simple similarity calculations, failing to dynamically adapt to changes in user behavior and lacking the ability to optimize ad narrative logic and multi-type ad sequences. Therefore, they exhibit significant shortcomings in multi-type ad delivery, ad sequence planning, and long-term user conversion value, easily resulting in low ad recommendation accuracy and impacting ad performance and user experience.

[0003] Existing technologies suffer from biased user behavior analysis, leading to low accuracy in ad recommendations and impacting ad performance and user experience. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for recommending multiple types of advertisements based on user profiles, in order to solve the technical problem that existing technologies suffer from one-sided user behavior analysis, resulting in low accuracy of advertisement recommendations and affecting the effectiveness of advertisement delivery and user experience.

[0005] In view of the above problems, this application provides a method and system for recommending multiple types of advertisements based on user profiles.

[0006] The first aspect of this application provides a method for recommending multiple types of advertisements based on user profiles. This method includes: analyzing the temporal dependencies between behaviors based on historical behavioral data of users across multiple interaction pathways, under a privacy protection mechanism; performing behavioral probability analysis based on the temporal dependencies between behaviors to generate a probabilistic user profile; generating an advertisement contact sequence arranged according to a preset narrative logic using a sequence decision model based on the probabilistic user profile and current context information; and matching corresponding advertisement content to user terminals according to the advertisement type and narrative advertisement requirements of each node in the sequence, and delivering and pushing the advertisements to user terminals according to the temporal relationship of the sequence.

[0007] Optionally, user click feedback data on the delivered advertising sequence is collected to obtain the feedback reading probability and purchase execution result; the probabilistic user profile and the sequence decision model are updated and optimized using the feedback reading probability and purchase execution result.

[0008] Optionally, based on the temporal dependencies between behaviors, the types of purchased goods, interaction paths and content formats, and temporal interaction relationships of user behaviors are extracted; the temporal dependencies are classified according to the types of purchased goods to construct a temporal dependency relationship for behaviors of multiple types of purchased goods; for each type of purchased goods, based on the interaction paths and content formats, and the temporal interaction relationships, the content formats and their sequential trigger probabilities in different interaction paths are identified to construct a user behavior transition probability map; based on the user behavior transition probability map for each type of purchased goods, the user's preference probability value and confidence interval for each type of purchased goods are calculated, and the probabilistic user profile is constructed using the preference probability value and confidence interval for each type of purchased goods.

[0009] Optionally, the nodes in the transition probability graph represent interaction behavior types or topic content, and the edge weights represent transition probabilities and average transition times. A recurrent neural network model is used to calculate and update in real time the user's preference probability for interaction content in each type of purchased goods and the confidence interval of the probability based on the traversal path of the nodes in the transition probability graph.

[0010] Optionally, the input user profile and context information are encoded into a state vector; at each step, an action is selected from the combined advertising action space, wherein the action is a triple including advertising type, narrative perspective, and delivery delay; based on the selected action, actions are recursively selected from the initial state through a policy network to form an action sequence, which is used as a blueprint for the advertising contact sequence; with the goal of maximizing long-term accumulated rewards, the sequence is optimized based on the blueprint of the advertising contact sequence, and a structured advertising contact sequence that meets the goal is output, and the model decision result is output, wherein the reward includes short-term advertising interaction metrics, estimated long-term conversion value, and a measure of sequence narrative fluency.

[0011] Optionally, a feedback monitoring point is set in the blueprint of the ad contact sequence. When the sequence is executed to the feedback monitoring point, feedback deviation information is obtained. Based on the feedback deviation information, if it is determined that the user interaction deviates from the expectation, real-time replanning of the subsequent sequence is triggered until the target is reached, and a structured ad contact sequence is output.

[0012] Optionally, the user's current multi-dimensional context information is obtained, including time information, geographical location information, device status information, and real-time behavior sequence information within the current session; the probabilistic user profile and the current context information are fused and input into the sequence decision model, which is instantiated into multiple sequentially connected advertising narrative units according to a preset narrative logic, and the connection time interval of the multiple advertising narrative units is planned based on the user behavior transition probability, and the advertising contact sequence is output.

[0013] Optionally, the narrative advertising requirements are parsed into quantifiable narrative advertising description dimensions, wherein the narrative advertising description dimensions include emotional tone, content form, content style, and information density; based on the narrative advertising description dimensions, a structured tag index is established for the advertising content library, and a matching library for each dimension of content elements is constructed, wherein the tag index of each material contains its quantified feature value in each narrative description dimension; using the advertising type and narrative advertising requirements to match in the matching library of each dimension of content elements, the content elements corresponding to the narrative advertising description dimensions are obtained, and advertising content is filtered.

[0014] Optionally, purchase path relationships are analyzed based on the temporal dependencies between behaviors and the user's purchase execution results to obtain path difference parameters. These path difference parameters are key influencing parameters for analyzing whether a user will complete a transaction on other platforms after being exposed to an advertisement on the target platform. The path difference parameters include price difference sensitivity, product information completeness difference, platform trust perception difference, and content uniqueness difference. Based on the path difference parameters, a cross-path isolation strategy is generated for the target transaction advertisement type to reduce path difference parameters and increase the transaction volume of the target transaction advertisement type.

[0015] A second aspect of this application provides a multi-type advertising recommendation system based on user profiles. This system includes: a behavior analysis module, used to analyze the temporal dependencies between behaviors based on historical behavior data of users across multiple interaction pathways, under a privacy protection mechanism; a probability analysis module, used to perform behavior probability analysis based on the temporal dependencies between behaviors, generating a probabilistic user profile; a contact sequence generation module, used to generate an advertising contact sequence based on the probabilistic user profile and current context information, through a sequence decision model, comprising a pre-defined narrative logic arrangement of multiple advertising types; and an advertising push module, used to match corresponding advertising content according to the advertising type and narrative advertising requirements of each node in the advertising contact sequence, and to push the content to user terminals according to the temporal relationship of the sequence.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application, based on historical user behavior data across multiple interaction pathways, analyzes the temporal dependencies between behaviors under a privacy protection mechanism; performs probabilistic behavior analysis based on these dependencies to generate a probabilistic user profile; and, based on the probabilistic user profile and current context information, generates an ad exposure sequence with multiple ad types arranged according to a preset narrative logic using a sequence decision model. Based on the ad exposure sequence, it matches corresponding ad content according to the ad type and narrative ad requirements of each node in the sequence, and pushes the ad to the user's terminal according to the temporal relationship of the sequence. This achieves the technical effect of generating personalized, temporally ordered ad sequences based on user profiles, enabling accurate recommendations of multiple ad types, and improving ad delivery effectiveness and user experience.

[0017] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the multi-type advertising recommendation method based on user profiles provided in this application.

[0020] Figure 2 A schematic diagram of the structure of the multi-type advertising recommendation system based on user profiles provided in this application.

[0021] Figure labeling: Behavior analysis module 11, probability analysis module 12, contact sequence generation module 13, advertising push module 14. Detailed Implementation

[0022] This application provides a method and system for recommending multiple types of advertisements based on user profiles. It addresses the technical problem that existing technologies suffer from biased user behavior analysis, leading to low accuracy in advertisement recommendations and negatively impacting advertising effectiveness and user experience. The system achieves the technical effect of generating personalized, time-series-based advertisement sequences based on user profiles, enabling accurate recommendations of multiple types of advertisements, and improving advertising effectiveness and user experience.

[0023] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0024] Example 1, as Figure 1 As shown, this application provides a method for recommending multiple types of advertisements based on user profiles, the method comprising: Based on users' historical behavior data across multiple interaction channels, and under the protection of privacy mechanisms, the temporal dependencies between behaviors are analyzed.

[0025] Specifically, based on the collection of user consent information, historical behavioral data of users across multiple interaction channels is collected through data interface APIs of various interaction channels. These interaction channels include user activity records in various online environments such as e-commerce platforms, social media, and search engines. Each behavioral data point includes user actions such as clicks, searches, browsing, comments, and purchases, and these actions have a temporal order and dependencies. To protect user privacy, de-identification processing is performed when collecting historical user behavioral data, such as encrypting or anonymizing user identity information to ensure the security and privacy of user data. For example, differential privacy is used by adding Laplace noise or Gaussian noise to the historical behavioral data to obscure the impact of individual user data on the overall data analysis results, thereby protecting user privacy. The noise level is set according to the sensitivity of the data and the strength of privacy protection.

[0026] Under this privacy protection mechanism, historical behavioral data generated by users across various interaction channels is encoded and sorted according to timestamps. Different behavior types, such as browsing, clicking, favoriting, and purchasing, are mapped to calculable numerical or vector features and sorted in ascending order based on precise timestamps, forming a personal behavioral time series for each user. For each type of behavioral time feature, autocorrelation analysis is used. By setting different time lag orders, the Pearson correlation coefficient or normalized covariance between the current time and the lag time is calculated to obtain the correlation coefficient of the time series under different time lag conditions. By comparing the significance of zero-order correlation and multi-lag correlation, repetitive or periodic patterns of user behavior over specific time intervals are identified.

[0027] Using the sliding window method, a window size of 5 days or 7 days and a step size of 1 day are set. Within each time window, the frequency of user behavior, the distribution of behavior types, and the combination of cross-channel behaviors are statistically analyzed. The changing trends of adjacent windows are compared to reflect the continuity and dependence of behavior over time. For example, in a continuous window, it can be analyzed that users usually browse advertising content in short videos, then compare the clicks on product details on multiple e-commerce platforms, and then add items to their cart. This can determine that the behavior chain has a significant sequential dependence in the time dimension.

[0028] By analyzing the temporal dependencies in user behavior data, we can more accurately understand user behavior patterns, improve the matching degree between advertisements and user needs, enhance the relevance and accuracy of ad recommendations, and thus improve the conversion rate of ad placement.

[0029] Based on the temporal dependencies between the behaviors, a probabilistic user profile is generated through behavioral probability analysis.

[0030] Furthermore, based on the temporal dependencies between the behaviors, a probabilistic user profile is generated through behavioral probability analysis, including: extracting the type of purchased goods, interaction path and content format, and temporal interaction relationship of user behavior based on the temporal dependencies between behaviors; classifying the temporal dependencies according to the purchased goods type to construct behavioral temporal dependencies for multiple purchased goods types; for each type of purchased goods, based on the interaction path and content format, and temporal interaction relationship, identifying the content format and sequential trigger probability of the user in different interaction paths, and constructing a user behavior transition probability map; calculating the user's preference probability value and confidence interval for each purchased goods type based on the user behavior transition probability map for each purchased goods type, and constructing the probabilistic user profile using the preference probability value and confidence interval for each purchased goods type.

[0031] Furthermore, in the transition probability graph, nodes represent interaction behavior types or topic content, and edge weights represent transition probabilities and average transition times. A recurrent neural network model is used to calculate and update in real time the user's preference probability for interaction content in each type of purchased goods and the confidence interval of the probability based on the traversal path of the nodes in the transition probability graph.

[0032] Specifically, based on the temporal dependencies between behaviors, behavioral features related to purchase decisions are extracted from the user's behavioral time series. These features include the types of goods purchased, the interaction channels and content formats involved, and the temporal interaction relationships between different behaviors. The types of goods purchased refer to the classification of potential transaction objects according to product categories or service categories, such as electronics, clothing, and food. The interaction channels refer to the environments in which the behavior occurs, such as e-commerce platforms, social media, and search engines. The content formats refer to information carriers such as images and text, short videos, live streams, and advertising materials. The temporal interaction relationships clarify the chronological order and time intervals of the behaviors. For example, a user might first see an image and text advertisement for a clothing brand on social media, then browse the brand's clothing details page on an e-commerce platform one day later, and finally complete the purchase. In this scenario, the purchased product type is clothing, the interaction channels are social media and the e-commerce platform, and the content formats are image and text advertisements and browsing the product details page.

[0033] Each piece of user behavior data is associated with its corresponding purchased product type. This product type can be determined based on the user's actual purchase history, add-to-cart history, or through behavior content parsing and product tag mapping. For example, behaviors involving comparing mobile phone parameters and browsing review videos are categorized under electronics, while behaviors involving outfit recommendations and size evaluations are categorized under clothing. Based on the temporal dependencies between behaviors, behaviors are labeled according to product type. Behavior sequences within the same product type are then split and classified from the temporal dependencies, forming multiple behavior time sequences based on product type. During the splitting process, the original timestamps and chronological order of the behaviors are preserved, ensuring that each product type sequence is an independent and complete temporal dependency analysis object, thus obtaining the temporal dependencies of behaviors across multiple purchased product types. This classification approach allows for more accurate analysis of behavioral characteristics under each product type, avoiding interference between different product decision-making logics.

[0034] For the temporal dependencies of each product type in a multi-product purchase category, this study combines interaction paths and content format characteristics to statistically identify the probability of subsequent actions triggered by users encountering various content formats through different interaction paths. A user behavior transition probability graph is constructed, employing conditional probability and Markov chains for probability analysis. Specifically, the temporal dependencies of a particular product type are discretized, representing behaviors such as browsing, clicking, searching, adding to favorites, adding to cart, and purchasing, along with their corresponding interaction paths and content formats, as a finite set of states. Using a first-order Markov chain model, assuming that the user's current behavioral state depends only on the previous behavioral state, the transition frequency between any two adjacent behavioral states is statistically analyzed by traversing temporal dependencies. Conditional probability calculations represent the probability of triggering the next behavioral state given the current behavioral state, interaction path, and content format. This transition probability is calculated by dividing the number of times a certain state transition occurs by the total number of times the current state occurs. This transition probability forms the Markov chain of user behavior for that product type, reflecting the probabilistic dependencies between different behaviors during the user's decision-making process for that product type. For each product type, the states in the Markov chain are further refined into combination nodes of behavioral type, interaction path, and content format, such as social media-image / text-browse or e-commerce platform-video-click. The user's historical behavioral paths for that product type are mapped to directed paths between nodes. All historical paths are statistically analyzed, and the number of transitions from any node to the next node is calculated. Based on this, the conditional transition probability is calculated and used as the probability weight of the edges in the transition probability graph. Simultaneously, by recording the time interval of each behavior transition, the average transition time between the same node pairs is calculated, and this average value is used as the time weight of the edges to form a user behavior transition probability graph. In this graph, nodes represent specific interaction behavior types or themes, and edges represent the transition relationship from one behavior or content to another. The edge weights include both the probability of the transition and the average transition time, reflecting the intensity and timeliness of the behavior transition. For example, when purchasing clothing, after seeing an image / text ad on social media, there is a 0.3 probability that a user will immediately search for that brand's clothing on an e-commerce platform, and a 0.2 probability that they will check related reviews on social media again after a period of time. These probabilities constitute the edge weights of the behavior transition probability graph. Furthermore, the edge weights can also include the average transition time, such as an average time of 2 hours from seeing the image / text ad to searching on the e-commerce platform.

[0035] The user behavior transition probability map constructed under each purchased product type is mapped to a sequence input. A recurrent neural network model is used to calculate the user's preference probability value for each purchased product type and the confidence interval of the probability. The recurrent neural network model adopts a two-layer long short-term memory network architecture, where each layer contains 128 hidden units to capture long-range dependencies and temporal dynamic relationships in the behavior sequence. The input layer encodes the features of each node into a vector representation, including behavior type, interaction path, content form, and the node's embedding vector in the transition probability map. The vector dimension is 64 to preserve the semantic and structural information between behaviors. The first layer LSTM captures the local temporal dependencies and short-term behavior patterns in the sequence. The input is the feature vector of each node in the sequence, and the output is the hidden state vector corresponding to each time step, with a length of 128, containing the encoded information of the sequence context. The LSTM unit contains an input gate, a forget gate, and an output gate. The input gate controls the writing of the current node's feature information, the forget gate controls the proportion of the previous state information retained, and the output gate controls the output of the hidden state to the next layer. The second LSTM layer, building upon the first layer's encoding, captures long-range dependencies and learns high-level patterns across time steps, such as relationships between user behavior over multiple days and platforms. The input is a 128-dimensional hidden state sequence from the first LSTM layer, and the output is a 128-dimensional hidden state from the second layer at each time step, which serves as input to the fully connected layer. The second LSTM layer combines the local dependency information captured by the first layer into a global temporal pattern. The output layer is a fully connected layer that uses the Softmax activation function to map the hidden states of the second LSTM layer to preference probabilities triggered by the next node's behavior, ranging from 0 to 1. This represents the likelihood of the user choosing that particular product type. For example, for electronic products, an output probability of 0.7 means the user has a 70% chance of choosing that type of product. Simultaneously, combining historical behavior transition frequency and path traversal counts, a number of subsamples are randomly drawn with replacement from the original path samples using Bootstrap sampling, repeated 1,000 times. The corresponding node transition probability is calculated for each sampling, and the probability value is recorded each time, forming a probability sample distribution. Percentiles are calculated based on this probability sample distribution; the 2.5th and 97.5th percentiles are used as the lower and upper bounds of the 95% confidence interval to estimate the 95% confidence interval for the predicted probability, reflecting the uncertainty of the probability estimation. The width of the confidence interval reflects the reliability of the probability estimation; the narrower the width, the more accurate the estimation and the higher the reliability. During training, the cross-entropy loss function is used as the objective function, and the Adam optimizer iteratively updates the parameters to learn the user's path patterns on the behavior transition graph. The training data includes the user's historical behavior paths and corresponding preference probability values ​​and confidence intervals for each purchased product type. The training batch size is set to 64, and the learning rate is 0.001. Iteration continues until the cross-entropy on the validation set converges or the early stopping strategy is triggered.

[0036] By using the preference probability of each type of purchased goods as the central value of the user's preference for that type, and the confidence interval as the reliable range of the probability, a vectorized or distributed probabilistic user profile is formed through the joint representation of multiple product types. This probabilistic user profile can accurately reflect the user's interests, preferences, and behavioral patterns. By transforming the complex behavioral paths of users in the transition probability graph into quantifiable probabilistic preference information and confidence intervals, and capturing the long-term dependencies of behavioral sequences through a recurrent neural network, the probabilistic user profile can be updated in real time with the user's latest behavior. This not only ensures the personalization and dynamism of advertising recommendations but also provides stable and predictable user state input for the generation of subsequent advertising contact sequences. This overcomes the static and one-sided nature of traditional user profiles, improves the matching degree between advertisements and user needs, enhances the relevance and accuracy of advertising recommendations, and thus significantly improves the click-through rate and conversion rate of advertising, thereby improving the effectiveness of advertising.

[0037] Based on the probabilistic user profile and the current context information, an ad contact sequence is generated by using a sequence decision model, which arranges multiple ad types according to a preset narrative logic.

[0038] Furthermore, the sequence decision model is trained based on a meta-reinforcement learning framework. The execution process of the sequence decision model includes: encoding the input user profile and context information into a state vector; selecting an action from the combined advertising action space at each step, wherein the action is a triple including advertising type, narrative angle, and delivery delay; recursively selecting actions from the initial state based on the selected action, forming an action sequence, and using the action sequence as a blueprint for the advertising contact sequence; optimizing the sequence based on the blueprint of the advertising contact sequence with the training objective of maximizing long-term accumulated rewards, outputting a structured advertising contact sequence that meets the objective, and outputting the model decision result, wherein the reward includes short-term advertising interaction metrics, estimated long-term conversion value, and a measure of sequence narrative fluency.

[0039] Specifically, a sequence decision model is trained based on a meta-reinforcement learning framework. First, each user task is defined as an environment instance, with its state composed of the user's probabilistic profile and current context information. The advertising action space consists of triples representing ad type, narrative perspective, and delivery delay. During training, the sequence decision model recursively generates a sequence of advertising actions in each task through a policy network. It calculates cumulative rewards based on short-term interaction metrics, long-term conversion value, and narrative fluency after the sequence execution. The policy network is used to select the optimal action given a state. The input to the policy network is the current environment state vector, containing the user's probabilistic profile and current context information. The input vector is first mapped to a high-dimensional feature representation through a fully connected embedding layer to preserve the combined features of the user profile and context. Then, it is input to a gated recurrent unit layer to ensure the sequence decision model can understand the continuity of the advertising narrative and the long-term correlation of user behavior patterns. The output passes through a fully connected output layer and is mapped to an action space probability distribution, where each action corresponds to the probability of the triples representing ad type, narrative perspective, and delivery delay. In the meta-learning phase, the policy network parameters are updated using a policy gradient algorithm with cumulative reward (a comprehensive score of short-term interaction metrics, long-term conversion value, and sequential narrative fluency) as the objective function. Simultaneously, in the meta-reinforcement learning phase, shared policies for multi-user, multi-context tasks are learned for initialization, enabling the sequence decision model to quickly adapt to new environments with minimal gradient updates when encountering new users or scenarios. Through multi-task training and meta-policy optimization, the resulting sequence decision model can transfer between different user tasks and generate ad contact sequences that maximize long-term cumulative rewards, achieving personalized and dynamic ad sequence recommendations.

[0040] The execution process of the sequential decision-making model includes: vectorizing the probabilistic user profile, using the preference probability and corresponding confidence interval of each product type as an independent feature dimension (e.g., forming a 2n-dimensional vector for n product types); encoding the context information; and mapping the user profile vector and context vector to a unified-dimensional state vector through a fully connected embedding layer, which serves as the environment state input for the reinforcement learning model. This process simultaneously preserves user interest and preference information and real-time context features, providing a complete state representation for the policy network to generate personalized and dynamic advertising actions. At each step, an action is selected from the combined advertising action space, which is a set containing all possible action combinations. Each action is represented as a triple, including the ad type, narrative perspective, and delivery delay. The ad type includes various forms, such as image ads, video ads, and text ads, each suitable for different scenarios and user groups. The narrative perspective refers to the way the ad content is presented, such as whether it focuses on the product's functional features or the user's emotional needs. The delivery delay indicates how long after the current moment the ad will be shown to the user, describing the executable strategy selection at each step.

[0041] The sequential decision model, through a policy network, recursively selects actions in the initial state. Based on historical user behavior preferences, transition probabilities, and contextual features, it generates a complete action sequence and uses this sequence as a blueprint for the ad exposure sequence, planning the specific form and content of ads that users will encounter at different times. For example, in the initial state where a user has just opened a shopping app, the first action selected is a video ad, narrating from the perspective of product functionality, with a 0-second delay—that is, an immediate video ad telling a story from the perspective of product functionality is displayed. The second action, based on the new state, uses the policy network again to select an action, such as selecting an image ad, narrating from the perspective of user emotional needs, with a 15-second delay—that is, an image ad telling a story from the perspective of user emotional needs is displayed 15 seconds later. This process continues, gradually forming a complete action sequence. With the goal of maximizing long-term accumulated rewards, sequence optimization is performed based on a blueprint of ad exposure sequences. The rewards include short-term ad interaction metrics, estimated long-term conversion value, and a measure of sequence narrative fluency. Short-term ad interaction metrics, such as click-through rate and viewing time, directly reflect the user's immediate interest in the ad. Estimated long-term conversion value refers to predicting the likelihood and amount of purchase behavior that a user will make due to exposure to the ad sequence in the future. The measure of sequence narrative fluency is used to evaluate the coherence and coordination between the various ads in the ad sequence in terms of narrative logic and style, to ensure that users have a good experience throughout the ad exposure process.

[0042] By continuously adjusting the parameters of the policy network using the policy gradient algorithm in reinforcement learning, the model maximizes long-term cumulative rewards during the selection of actions to form action sequences. For example, during training, the sequence decision model repeatedly tries different action sequences, calculates the long-term cumulative reward for each sequence, and then updates the parameters of the policy network based on the reward value, making it more likely to select action sequences that yield higher rewards in the future. Finally, the model outputs a structured ad contact sequence that meets the objective as its decision result. Each ad node in the structured ad contact sequence includes ad type, narrative perspective, and delivery timing, forming a personalized and logically consistent ad blueprint to guide e-commerce platforms in precise ad delivery.

[0043] For example, a user profile shows that the user is a young woman who frequently buys beauty and apparel products and recently browsed sneakers. The context information, such as it being a weekend afternoon and the user browsing the sports section, is encoded as a state vector. Actions are selected from the combined ad action space, such as selecting a sneaker ad, a first-person narrative, and immediate delivery (a triple). The policy network recursively selects actions from the initial state to form an action sequence, optimizing to maximize long-term accumulated rewards. If an ad has a high click-through rate in the short term, it is estimated to encourage the user to buy sneakers in the long term, and the narrative sequence is coherent and smooth, resulting in a higher reward. The final output is a model decision that first pushes a new sneaker ad, then an immediate delivery via a first-person narrative, followed by an ad for matching sportswear, and finally a third-person narrative delivered one hour later. This generates a personalized ad contact sequence blueprint, guiding the e-commerce platform to make precise ad placements and achieve dynamic ad sequence recommendations.

[0044] By transforming user profiles and contextual information into executable advertising strategies, the sequence decision model trained through meta-reinforcement learning can generate advertising contact sequences that satisfy the maximization of long-term benefits and optimization of narrative logic. This enables the transformation of advertising recommendation from static matching to dynamic, personalized, and sequential delivery, thereby improving user acceptance and engagement with advertising and ultimately enhancing advertising effectiveness and conversion rates.

[0045] Furthermore, based on the blueprint of the ad contact sequence, sequence optimization is performed to output a structured ad contact sequence that meets the target. This includes: setting feedback monitoring points in the blueprint of the ad contact sequence; obtaining feedback deviation information when the sequence reaches the feedback monitoring point; and triggering real-time replanning of subsequent sequences based on the feedback deviation information when it is determined that the user interaction deviates from the expectation, until the target is reached, and outputting a structured ad contact sequence.

[0046] Specifically, the ad engagement sequence generated by the sequential decision model is used as a blueprint, where each node represents an ad delivery and narrative unit, including ad type, narrative requirements, and planned delivery time. Feedback monitoring points are set in the blueprint of the ad engagement sequence, typically located at key nodes or every few nodes. When the sequence reaches a feedback monitoring point, user interaction data is collected, such as click-through rate (CTR), browsing duration, interactive behavior, and purchase conversion. By comparing the actual feedback with expected metrics, such as predicted click probability or conversion probability, feedback deviation information is calculated. Feedback deviation information reflects the difference between the actual user interaction and the pre-set expected target, quantifying the degree to which user behavior deviates from expectations. For example, if the expected CTR for an ad at a certain node is 0.6, but the actual CTR is only 0.35, the deviation value is 0.25.

[0047] A threshold judgment method is adopted, setting a deviation range threshold, such as ±0.5. When the feedback deviation information exceeds the deviation range threshold, it is determined that the user interaction deviates from the expectation, and a real-time replanning of the subsequent sequence is triggered. Using reinforcement learning algorithms, based on the current feedback deviation information, the ad type, narrative content and delivery time interval of the subsequent nodes are adjusted through strategy updates to rematch user interests and behavioral trends. This ensures that the final ad contact sequence can still maximize long-term cumulative rewards, such as short-term interaction indicators, long-term conversion value and sequence narrative fluency, and output a structured ad contact sequence. This structured ad contact sequence has been optimized and adjusted multiple times to better adapt to the actual needs and behavioral patterns of users.

[0048] By dynamically adjusting the ad delivery sequence based on real-time user feedback, ad targeting becomes more precise and flexible. Continuous optimization of the sequence effectively enhances ad appeal and relevance, thereby increasing user engagement and ad performance, improving conversion rates and user experience, and ensuring sustained optimization of ad performance.

[0049] Furthermore, based on the probabilistic user profile and the current context information, an ad contact sequence is generated by a sequence decision model, which arranges multiple ad types according to a preset narrative logic. This includes: obtaining the user's current multi-dimensional context information, including time information, geographical location information, device status information, and real-time behavior sequence information within the current session; fusing the probabilistic user profile and the current context information and inputting them into the sequence decision model; the sequence decision model instantiates multiple sequentially connected ad narrative units according to the preset narrative logic; and plans the connection time interval of the multiple ad narrative units based on the user behavior transition probability, and outputs the ad contact sequence.

[0050] Specifically, the system acquires multi-dimensional contextual information about the user in real time, including time information, geographic location information, device status information, and real-time behavioral sequence information within the current session. Time information reflects the user's current time context, such as specific time, day of the week, or month, and can be represented by hourly and weekday periodic encoding. Geographic location information reflects the user's region or consumption environment, such as city-level location or business district tags, and is represented by location embedding vectors. Device status information describes the conditions of the user's current terminal, such as mobile / PC, network status, and screen size. Different devices offer different advertising effects and user operation convenience. Real-time behavioral sequence information within the current session reflects short-term changes in the user's interests during the session, such as the sequence embedding of browsing pages, clicking on products, and searching keywords. The probabilistic user profile and the current contextual information are then fused. The probabilistic user profile uses the preference probability and confidence interval of each purchased product type as core features to represent the user's stable interests and uncertainties in the long term. The fusion method can involve feature concatenation followed by mapping to a unified-dimensional state vector through a fully connected embedding layer, which serves as the input state for the sequential decision model.

[0051] The sequential decision-making model, based on a pre-defined narrative logic—the progressive relationship between advertising content—from awareness and interest to comparison and conversion, instantiates the entire advertising delivery process into multiple sequentially connected advertising narrative units. Each narrative unit corresponds to an advertising type and narrative perspective, such as brand awareness, feature explanation, or promotional conversion. Simultaneously, the sequential decision-making model incorporates user behavior transition probabilities, such as the probability of transitioning from content browsing to product search and the average transition time, to plan the delivery time interval between adjacent narrative units. This forms an advertising exposure sequence with time-controlled rhythm. The user behavior transition probability is obtained by analyzing a large amount of historical user behavior data, reflecting the likelihood that a user will take the next action, such as clicking, purchasing, or skipping, within a certain time after seeing an advertisement. For example, if a probabilistic user profile shows that a user's preference for smartphones is 0.72, and the average transition time from review videos to product searches is 3 hours in a nighttime mobile scenario, review video ads can be pushed first, followed by product comparison or promotional ads after an interval of about 3 hours. This ultimately outputs an ad contact sequence that conforms to the evolution of user interests and narrative coherence, improving the user's ad experience and acceptance, thereby helping to increase ad click-through rates and conversion rates.

[0052] By combining probabilistic user profiles of long-term user preferences with contextual information from short-term real-time situations, and upgrading single ad recommendations into ad exposure sequences with narrative structure and temporal rhythm through a sequential decision model, the system avoids the fragmented experience caused by fragmented and randomized ad delivery. By introducing behavioral transfer probability and narrative logic constraints, the system not only improves the coherence of ad content and user acceptance, but also helps to gradually guide users through the decision-making path from cognition to conversion in multiple exposures, ultimately improving overall ad conversion efficiency and long-term value.

[0053] Based on the ad contact sequence, the corresponding ad content is matched according to the ad type and narrative ad requirements of each node in the sequence, and then pushed to user terminals according to the time relationship of the sequence.

[0054] Furthermore, according to the ad type and narrative ad requirements of each node in the sequence, the corresponding ad content is matched, including: parsing the narrative ad requirements into quantifiable narrative ad description dimensions, wherein the narrative ad description dimensions include emotional tone, content form, content style, and information density; based on the narrative ad description dimensions, a structured tag index is established for the ad content library, and a matching library for content elements of each dimension is constructed, wherein the tag index of each material contains its quantified feature value on each narrative description dimension; using the ad type and narrative ad requirements to match in the matching library of content elements of each dimension, the content elements corresponding to the narrative ad description dimensions are obtained, and ad content is filtered.

[0055] Specifically, using the advertising contact sequence output by the sequential decision model as the execution blueprint, each node in the advertising contact sequence is regarded as an advertising narrative unit with a clear objective. Each narrative unit contains the corresponding advertising type and narrative advertising requirements. The narrative advertising requirements are parsed and transformed into quantifiable narrative advertising description dimensions, including emotional tone, content form, content style, and information density. Among these, emotional tone describes the emotional tendency of the advertisement, such as positive and enthusiastic, warm and touching, humorous and witty, etc. Semantic analysis of the advertising copy text can be performed using sentiment analysis algorithms in natural language processing to identify words and sentence structures that express positive, negative, or neutral emotions. By statistically analyzing the proportion of positive and negative words, these are quantified into numerical values ​​between 0 and 1 to calculate the sentiment score. The sentiment score expression is as follows: S=(N p -N n ) / N t And normalize S 归 =(S+1) / 2, where N p N n N tThese represent the number of positive words, the number of negative words, and the total number of words, respectively. A sentiment score closer to 1 indicates a more positive sentiment, 0.5 indicates neutrality, and a score closer to 0 indicates a more negative sentiment. Content formats include images, videos, and text, determined directly based on the physical attributes of the advertising material. Images, videos, and text each correspond to different category tags and can be represented using one-hot encoding, such as [1,0,0] for images, [0,1,0] for videos, and [0,0,1] for text. Different formats are suitable for different advertising types and narrative needs. Content style reflects visual and expressive style, such as technological, lifestyle, and youthful. Features are extracted from advertising visual elements, images, video footage, and text expression, compared with a predefined style vector library, and cosine similarity is calculated to determine the style category and assign a corresponding label. Alternatively, style vectors can be used to represent the similarity between the content and various styles. Information density measures the amount of information carried per unit of time, such as low-density storytelling and high-density parameter descriptions. It can be calculated based on the number of key information items N and the length T of the ad creative. The information density expression is D=N / T, which is then normalized using a min-max normalization method. By segmenting and analyzing the ad content, the number of key information items appearing per unit of time is counted, and combined with the ad duration, the information carrying capacity is calculated.

[0056] Based on the narrative advertising description dimension, each creative in the advertising content library undergoes multi-dimensional feature extraction in the manner described above. Each creative is then quantified in terms of its attributes within the narrative advertising description dimension. This advertising content library is a centralized database storing and managing various advertising-related creatives and information, containing diverse types, styles, and themes of advertising creatives to provide rich and varied content support for advertising placement and marketing activities. Creatives refer to the basic elements or original materials that constitute advertising content; they are the specific entities within the advertising content library. The quantified features are uniformly stored as a structured tag index. This involves generating a multi-dimensional vector or table record for each creative, containing emotional intensity values, content format encoding, style vectors, and information density scores. A matching library for content elements across each dimension is constructed. By organizing the feature values ​​of the same dimension by category or numerical range (e.g., emotional intensity 0-0.3 for low, 0.3-0.7 for medium, and 0.7-1 for high), it facilitates rapid retrieval and filtering of creatives matching the dimensional features according to the narrative advertising requirements during the subsequent advertising contact sequence matching stage, achieving efficient matching and refined placement of advertising content. The ad type and narrative ad requirements of the preceding sequence nodes are determined through multi-dimensional similarity calculations or rule filtering within a content element matching library. This process selects ad creatives that best match the narrative requirements in terms of emotional tone, content format, content style, and information density, thus achieving precise ad content selection. Common methods for multi-dimensional similarity calculation include weighted cosine similarity or Euclidean distance. Different weights can be assigned based on the importance of each dimension, such as 0.4 for emotional tone, 0.2 for content format, 0.3 for content style, and 0.1 for information density. Simultaneously, rule filtering conditions can be set to quickly eliminate creatives that do not meet basic requirements, such as requiring the content format to match video or the information density to be within a specified range. By using sorting or threshold filtering methods, the creatives with the highest similarity to the target feature vector are selected as the precise ad content for that sequence node, achieving a high degree of matching between ad content and narrative objectives, thereby ensuring the consistency of the ad narrative and the uniformity of the user experience.

[0057] Then, according to the planned time sequence and time intervals in the advertising contact sequence, the selected advertising content is delivered through the corresponding user terminals to achieve sequential and rhythmic advertising push. For example, in the advertising contact sequence for home appliances, the first node requires a low-information-density, emotionally neutral brand awareness advertisement, which can be matched with a short, lifestyle-oriented video clip from the content library. The second node requires a high-information-density, rational-style functional explanation advertisement, which can be matched with parameter explanation text or video, and pushed out after a preset 6 hours.

[0058] By quantifying the dimensions of narrative advertising descriptions and using structured indexing of the content library, we can achieve precise alignment between advertising content and narrative objectives. This avoids the problems of homogenized or emotionally fragmented continuous advertising, enabling advertising content to form a coherent narrative in time sequence. At the same time, it improves the efficiency of users' understanding and acceptance of advertising information, thereby enhancing the conversion effect and user experience of the overall advertising contact sequence.

[0059] Furthermore, after pushing ads to user terminals according to the sequential time relationship, the process also includes: collecting user click feedback data on the ad sequence to obtain feedback reading probability and purchase execution result; and using the feedback reading probability and purchase execution result to update and optimize the probabilistic user profile and the sequence decision model.

[0060] Specifically, after the ad exposure sequence is delivered to the user's terminal according to the sequence plan, user click feedback data on the ad sequence is collected in real time, and the feedback read probability and purchase execution result are obtained. The feedback read probability is used to quantify the user's actual attention to the ad content within a specific time window, and can be calculated by the ratio of clicks to impressions, i.e., feedback read probability = clicks / ad impressions. The purchase execution result directly reflects whether the user has completed the corresponding ad conversion action, such as placing an order or making a payment, and can be represented by a binary value: 1 indicates purchase, and 0 indicates no purchase.

[0061] The probability of receiving feedback and the purchase execution result are used as feedback signals for probabilistic user profile updates. A Bayesian update method is applied to the preference probability and confidence interval for each purchased product type: historical preference probabilities are treated as a prior distribution P(θ), and a likelihood function P(D|θ) is constructed from the newly observed feedback data. The updated posterior probability and its confidence interval are obtained using the Bayesian formula P(θ|D) = (P(D|θ)P(θ)) / P(D), reflecting the strength and uncertainty of the user's latest preferences. Simultaneously, for online optimization of the sequential decision model, user feedback is used as a reward signal input to the policy network. Incremental policy gradients are used to fine-tune the policy network parameters. The reward for each recursive action is a weighted combination of the user's click or purchase status for that node's advertisement. Gradient ascent updates the policy network, enabling the sequential decision model to learn the latest user behavioral preferences and improve subsequent advertising action selection. This allows the sequential decision model to more accurately predict user preferences and generate efficient advertising sequences. This process is executed in real time, enabling dynamic adaptability in ad sequence generation, while ensuring synchronous optimization of user profiles and sequence decision models after data updates, thereby improving click-through rates and conversion rates in the next round of ad campaigns.

[0062] By adjusting probabilistic user profiles and sequential decision models through real-time feedback, we can not only capture the dynamic changes in user interests, but also continuously optimize the ad sequence generation strategy, thereby improving ad click-through rates, conversion rates, and overall user experience, and achieving adaptive and iteratively optimized personalized ad recommendations.

[0063] Furthermore, the multi-type ad recommendation method based on user profiles also includes: analyzing the purchase path relationship based on the temporal dependency between behaviors and the user's purchase execution results to obtain path difference parameters. These path difference parameters are key influencing parameters for analyzing whether a user will flow to other platforms to complete a transaction after being exposed to an ad on a target platform. The path difference parameters include price difference sensitivity, product information completeness difference, platform trust perception difference, and content uniqueness difference. Based on the path difference parameters, a cross-path isolation strategy is generated for the target transaction ad type to reduce path difference parameters and increase the transaction volume of the target transaction ad type.

[0064] Specifically, a complete path from ad exposure to purchase execution is formed based on the temporal dependencies between behaviors. Each path includes the user's behavior type, interaction method, access time, and final purchase status. Combined with the user's purchase execution result, such as whether the order was placed on the target platform or the user moved to another platform, the purchase path relationship is analyzed to obtain path difference parameters. These path difference parameters are key influencing parameters for analyzing whether a user moves to another platform to complete a transaction after ad exposure on the target platform. They include price difference sensitivity, product information completeness difference, platform trust perception difference, and content uniqueness difference.

[0065] The price difference sensitivity is used to measure how sensitive users are to choosing a platform to purchase from due to price differences between different platforms. The price difference is obtained by calculating the difference between the price of the target platform and the prices of other platforms for each path. A logistic regression model is established based on the price difference, with whether the user completes the purchase on the target platform Y∈{0,1} as the dependent variable and the price difference ΔP as the independent variable: logit(P(Y=1))=α+β×ΔP, where α represents the basic probability that the user will purchase on the target platform when the price difference is 0, and the regression coefficient β is used as the price difference sensitivity index. The price difference sensitivity is obtained by solving the maximum likelihood estimation. β>0 indicates that the larger the price difference, the more likely the user is to flow to other platforms, and β<0 indicates that the user is not sensitive to price differences. The product information completeness difference is used to measure the impact of product information completeness on the user's purchase path on different platforms. It is achieved by quantitatively scoring the product information of the target platform and other platforms, such as product description, parameter specifications, and number of reviews, and then weighting the scores. The weight values ​​can be set according to actual needs to obtain the platform information completeness score. The information completeness difference ΔI between the target platform and other platforms is calculated, and statistical analysis or regression modeling is performed on the user's purchase path: P(Y=1)=f(ΔI) to obtain the product information completeness difference influence coefficient. The product information completeness difference influence coefficient value represents the sensitivity of the information completeness difference to the user's choice of target platform. Platform trust perception difference is used to measure the impact of users' varying trust levels across different platforms on purchasing behavior. It collects indicators such as users' historical ratings, platform credit scores, and brand loyalty. After normalizing these indicators, a weighted average is calculated to determine the trust level between the target platform and other platforms. The weights can be set based on actual needs and experience. The trust difference ΔT is calculated using the trust levels of the target platform and other platforms. This is then used to analyze and model purchasing behavior, where P(Y=1)=σ(γ0+γ1×ΔT), and σ is the Sigmoid function σ(x)=1 / (1+e^(-γ0+γ1×ΔT)). -x The linear prediction value is mapped to a probability range of 0-1. γ0 represents the base purchase probability when the trust difference is 0, and γ1 represents the user's sensitivity to the platform's trust difference. The larger the value, the more significant the impact of the trust difference on the purchase decision. Content uniqueness difference measures the impact of the target platform's exclusive content or experience differing from other platforms on the purchase path. This is achieved by calculating the uniqueness index of the target platform's advertising content or promotional information compared to other platforms for each path, such as the number of exclusive products, exclusive copywriting, or video materials, and then normalizing these indices. The content uniqueness difference ΔU is calculated, and a regression model P(Y=1)=f(ΔU) is established. The coefficient value ΔU serves as the content uniqueness difference parameter, used to determine the contribution of exclusive content to the target platform's sales.

[0066] By integrating parameters such as price difference sensitivity, product information completeness difference, perceived platform trust difference, and content exclusivity difference, path difference parameters are formed for each user path. Based on these parameters, cross-path isolation strategies are generated according to the target transaction ad type: If price difference sensitivity is the key parameter, the cross-path isolation strategy can embed exclusive limited-time price anchors or bundled offers in the target platform ad and emphasize their exclusivity in the copy. If content exclusivity is the key parameter, the cross-path isolation strategy can deliver customized narrative content, exclusive product previews, or in-depth brand interactive experiences. If information completeness difference is dominant, detailed product parameters, reviews, and user feedback summaries are provided to ensure information integrity. If trust difference is dominant, platform certifications, brand guarantees, and after-sales service commitments are displayed to create a content barrier that other platforms cannot access, thereby reducing the likelihood of users migrating to other platforms. For example, in an advertising recommendation project on an e-commerce platform, analysis of the purchase path revealed that price difference sensitivity was the main path difference parameter leading users to other platforms. Consequently, the platform embedded exclusive limited-time price anchors in the advertising content, reducing the price of some products by 5% and emphasizing in the copy that the discount was exclusive to this platform. After implementing this strategy, the transaction volume of the target advertising type increased by 25%, and the total transaction volume of the target platform increased by 18%, effectively improving the transaction volume of the target platform.

[0067] In a multi-platform environment, users' purchase paths may change. By quantitatively analyzing users' cross-platform purchase behavior and key influencing factors, and combining this with customized intervention strategies, a cross-path isolation strategy can be constructed. This strategy can dynamically adjust advertising content and delivery methods based on the differences in parameters across different paths, reducing user churn across platforms. It not only optimizes the conversion effect of ads on the target platform, but also enhances the exclusivity and attractiveness of advertising content, thereby effectively increasing the sales volume of the target advertising type.

[0068] Example 2, based on the same inventive concept as the multi-type advertising recommendation method based on user profiles in the previous examples, such as... Figure 2 As shown, this application provides a multi-type advertising recommendation system based on user profiles, wherein the multi-type advertising recommendation system based on user profiles includes: The behavior analysis module 11 is used to analyze the temporal dependencies between behaviors based on the user's historical behavior data across multiple interaction paths, under a privacy protection mechanism; the probability analysis module 12 is used to perform behavior probability analysis based on the temporal dependencies between behaviors, generating a probabilistic user profile; the contact sequence generation module 13 is used to generate an ad contact sequence with multiple ad types arranged according to a preset narrative logic based on the probabilistic user profile and current context information through a sequence decision model; the ad push module 14 is used to match the corresponding ad content according to the ad type and narrative ad requirements of each node in the ad contact sequence, and push the ad to the user terminal according to the temporal relationship of the sequence.

[0069] Furthermore, the system is also used to: collect user click feedback data on the sequence of advertisements, obtain feedback reading probability and purchase execution result; and use the feedback reading probability and purchase execution result to update and optimize the probabilistic user profile and the sequence decision model.

[0070] Furthermore, the probability analysis module 12 is also used to: extract the purchase product type, interaction path and content form, and temporal interaction relationship of user behavior based on the temporal dependency relationship between behaviors; classify the temporal dependency relationship according to the purchase product type to construct the behavioral temporal dependency relationship of multiple purchase product types; for each purchase product type, based on the interaction path and content form, and temporal interaction relationship, identify the content form and the probability of sequential triggering of the user in different interaction paths, and construct a user behavior transition probability map; calculate the user's preference probability value and confidence interval for each purchase product type according to the user behavior transition probability map of each purchase product type, and construct the probabilistic user profile using the preference probability value and confidence interval of each purchase product type.

[0071] Furthermore, the probability analysis module 12 is also used to: the nodes in the transition probability graph represent interaction behavior types or topic content, the edge weights represent transition probabilities and average transition times, and a recurrent neural network model is used to calculate and update in real time the user's preference probability for interaction content in each type of purchased goods and the confidence interval of the probability based on the traversal path of the nodes in the transition probability graph.

[0072] Furthermore, the contact sequence generation module 13 is also used to: encode the input user profile and context information into a state vector; select an action from the combined advertising action space at each step, the action being a triple including advertising type, narrative angle, and delivery delay; based on the selected action, recursively select actions from the initial state through a policy network to form an action sequence, and use the action sequence as a blueprint for the advertising contact sequence; with the goal of maximizing long-term accumulated rewards, perform sequence optimization based on the blueprint of the advertising contact sequence, output a structured advertising contact sequence that meets the goal, and output the model decision result, wherein the rewards include short-term advertising interaction metrics, estimated long-term conversion value, and a measure of sequence narrative fluency.

[0073] Furthermore, the contact sequence generation module 13 is also used to: set feedback monitoring points in the blueprint of the advertising contact sequence; when the sequence is executed to the feedback monitoring point, obtain feedback deviation information; and when it is determined that the user interaction deviates from the expectation based on the feedback deviation information, trigger real-time replanning of the subsequent sequence until the target is reached, and output a structured advertising contact sequence.

[0074] Furthermore, the contact sequence generation module 13 is also used to: obtain the user's current multi-dimensional context information, including time information, geographical location information, device status information, and real-time behavior sequence information within the current session; fuse the probabilistic user profile and the current context information, input them into the sequence decision model, the sequence decision model is instantiated into multiple sequentially connected advertising narrative units according to a preset narrative logic, and the connection time interval of the multiple advertising narrative units is planned based on the user behavior transition probability, and the advertising contact sequence is output.

[0075] Furthermore, the ad push module 14 is also used to: parse the narrative ad requirements into quantifiable narrative ad description dimensions, wherein the narrative ad description dimensions include emotional tone, content form, content style, and information density; based on the narrative ad description dimensions, establish a structured tag index for the ad content library, construct a matching library for content elements of each dimension, wherein the tag index of each material contains its quantified feature value in each narrative description dimension; and use the ad type and narrative ad requirements to match in the matching library of content elements in each dimension to obtain content elements corresponding to the narrative ad description dimension, and perform ad content filtering.

[0076] Furthermore, the system is also used to: analyze the purchase path relationship based on the temporal dependency between behaviors and the user's purchase execution results to obtain path difference parameters. These path difference parameters are key influencing parameters for analyzing whether a user will flow to other platforms to complete a transaction after being exposed to an advertisement on a target platform. The path difference parameters include price difference sensitivity, product information completeness difference, platform trust perception difference, and content uniqueness difference. Based on the path difference parameters, a cross-path isolation strategy is generated for the target transaction advertisement type to reduce path difference parameters and increase the transaction volume of the target transaction advertisement type.

[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The multi-type advertising recommendation method and specific examples based on user profiles in the foregoing embodiment one are also applicable to the multi-type advertising recommendation system based on user profiles in this embodiment. Through the foregoing detailed description of the multi-type advertising recommendation method based on user profiles, those skilled in the art can clearly understand the multi-type advertising recommendation system based on user profiles in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0079] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A multi-type advertisement recommendation method based on user profiles, characterized in that, include: Based on users' historical behavior data across multiple interaction channels, and under the protection of privacy mechanisms, the temporal dependencies between behaviors are analyzed. Based on the temporal dependencies between the behaviors, a probabilistic user profile is generated through behavioral probability analysis. Based on the probabilistic user profile and the current context information, an ad contact sequence consisting of multiple ad types arranged according to a preset narrative logic is generated through a sequence decision model; Based on the ad contact sequence, the corresponding ad content is matched according to the ad type and narrative ad requirements of each node in the sequence, and then pushed to user terminals according to the time relationship of the sequence.

2. The multi-type advertisement recommendation method based on user profiles according to claim 1, characterized in that, After delivering push notifications to user terminals according to the sequential time relationship, it also includes: Collect user click feedback data on the delivered ad sequence to obtain feedback read probability and purchase execution results; The probabilistic user profile and the sequential decision model are updated and optimized using the feedback reading probability and purchase execution result.

3. The method for recommending multiple types of advertisements based on user profiles according to claim 1, characterized in that, Based on the temporal dependencies between the behaviors, a probabilistic user profile is generated through behavioral probability analysis, including: Based on the temporal dependencies between behaviors, extract the types of goods purchased, interaction paths and content formats, and temporal interaction relationships of user behaviors; Classify the temporal dependencies according to the types of purchased goods, and construct the behavioral temporal dependencies for multiple types of purchased goods. For each type of shopping product, based on the temporal dependency relationship, and according to the interaction path, content form, and temporal interaction relationship, identify the content forms and sequential trigger probabilities of users in different interaction paths, and construct a user behavior transition probability map. Based on the user behavior transition probability map for each type of shopping product, calculate the user's preference probability value and confidence interval for each type of shopping product, and construct the probabilistic user profile using the preference probability value and confidence interval for each type of shopping product.

4. The multi-type advertisement recommendation method based on user profiles according to claim 3, characterized in that, The nodes in the transition probability graph represent interaction behavior types or topic content, and the edge weights represent transition probabilities and average transition times. A recurrent neural network model is used to calculate and update in real time the user's preference probability for interaction content in each type of purchased goods and the confidence interval of the probability based on the traversal path of the nodes in the transition probability graph.

5. The multi-type advertisement recommendation method based on user profiles according to claim 1, characterized in that, The sequence decision model is trained based on a meta-reinforcement learning framework, and the execution process of the sequence decision model includes: Encode the input user profile and context information into a state vector; At each step, select an action from the combined advertising action space, where the action is a triple including ad type, narrative perspective, and delivery delay; Based on the selected action from the initial state, actions are recursively selected through a policy network to form an action sequence, which is then used as a blueprint for an ad contact sequence. With the goal of maximizing long-term accumulated rewards, the algorithm optimizes the sequence based on the blueprint of the ad contact sequence, outputs a structured ad contact sequence that meets the goal, and outputs the model decision results. The rewards include short-term ad interaction metrics, estimated long-term conversion value, and a measure of sequence narrative fluency.

6. The multi-type advertisement recommendation method based on user profiles according to claim 5, characterized in that, Sequence optimization is performed based on the blueprint of the ad contact sequence to output a structured ad contact sequence that meets the objectives, including: Feedback monitoring points are set in the blueprint of the advertising contact sequence. When the sequence is executed to the feedback monitoring point, feedback deviation information is obtained. When the user interaction deviates from the expected behavior based on the feedback deviation information, real-time replanning of the subsequent sequence is triggered until the target is reached, and a structured ad contact sequence is output.

7. The multi-type advertisement recommendation method based on user profiles according to claim 6, characterized in that, Based on the probabilistic user profile and current context information, an ad contact sequence is generated using a sequence decision model, which arranges multiple ad types according to a preset narrative logic, including: Obtain the user's current multi-dimensional context information, including time information, geographical location information, device status information, and real-time behavior sequence information within the current session; The probabilistic user profile and current context information are fused and input into the sequence decision model. The sequence decision model is instantiated into multiple sequentially connected advertising narrative units according to a preset narrative logic, and the connection time interval of the multiple advertising narrative units is planned based on the user behavior transition probability, and the advertising contact sequence is output.

8. The multi-type advertisement recommendation method based on user profiles according to claim 1, characterized in that, Match the corresponding ad content according to the ad type and narrative ad requirements of each node in the sequence, including: Narrative advertising requirements are analyzed into quantifiable narrative advertising description dimensions, which include emotional tone, content form, content style, and information density. Based on the narrative advertising description dimensions, a structured tag index is established for the advertising content library, and a matching library for content elements in each dimension is constructed. The tag index of each material contains its quantitative feature value in each narrative description dimension. By matching the ad type and narrative ad requirements in the matching library of content elements in each dimension, the content elements of the corresponding narrative ad description dimension are obtained, and the ad content is filtered.

9. The method for recommending multiple types of advertisements based on user profiles according to claim 1, characterized in that, Also includes: Based on the temporal dependencies between behaviors and the purchase execution results of users, purchase path relationship analysis is performed to obtain path difference parameters. The path difference parameters are key influencing parameters for analyzing users' flow to other platforms to complete transactions after being exposed to advertisements on the target platform. The path difference parameters include price difference sensitivity, product information completeness difference, platform trust perception difference, and content uniqueness difference. Based on the path difference parameters, a cross-path isolation strategy is generated for the target transaction ad type to reduce the path difference parameters and increase the transaction volume of the target transaction ad type.

10. A multi-type advertising recommendation system based on user profiles, characterized in that, The step of implementing the multi-type advertising recommendation method based on user profiles according to any one of claims 1 to 9, wherein the multi-type advertising recommendation system based on user profiles comprises: The behavior analysis module is used to analyze the temporal dependencies between behaviors based on users' historical behavior data across multiple interaction channels, while maintaining a privacy protection mechanism. The probability analysis module is used to perform behavioral probability analysis based on the temporal dependencies between the behaviors and generate probabilistic user profiles. The contact sequence generation module is used to generate an ad contact sequence with multiple ad types arranged according to a preset narrative logic based on the probabilistic user profile and the current context information through a sequence decision model. The ad push module is used to match the corresponding ad content based on the ad contact sequence, according to the ad type and narrative ad requirements of each node in the sequence, and to push the ad to the user terminal according to the time relationship of the sequence.

Images