Artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis

Abstract

The invention provides an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis, and the method comprises the steps: constructing second-level, minute-level, hour-level and other behavior feature views of different time granularities for extracted original features, and forming multi-time scale feature representation; analyzing feature views under different time scales through a graph mining algorithm, mining an association rule between feature nodes, and constructing a cross-scale feature association graph; in the feature correlation map, according to the spectral clustering result of the feature nodes and the gravitation direction of the feature edges, the temporal correlation degree of different feature combinations is judged; if the time correlation degree of the feature combination exceeds an empirical threshold value, endowing the time weight of the feature combination with a high weight coefficient of power law distribution; and dynamically adjusting the network depth and the connection density of the association atlas according to the vertical dimension and the semantic hierarchy of the features, and optimizing the feature weight distribution in the atlas.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular, to an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis. Background Art

[0002] In advertisement pushing decisions, how to effectively quantify the contribution degrees of data in different time dimensions is a complex technical problem. User behavior data has multi-level time characteristics, such as long-term browsing preferences, short-term search records, seasonal periodic behaviors, etc. The influence of these data on advertisement pushing at different time scales is significantly different. In order to accurately capture these differences, a weight calculation model based on an attention mechanism needs to be constructed, which can dynamically evaluate the contribution degree of each time dimension to the current push. However, due to the time decay effect of data, the influence of historical data in different time dimensions on the current decision is not linear, but gradually weakens over time. Therefore, it is necessary to introduce a time decay factor to more precisely reflect the timeliness of historical data.

[0003] In addition, in advertisement pushing decisions, it is necessary to quantify the contribution degrees of data in different historical dimensions. For example, for data layers such as user long-term browsing preferences, short-term search records, and seasonal periodic behaviors, a contribution evaluation model based on a time decay factor is established to dynamically adjust the parameter ratio in feature engineering. A multi-granularity time analysis framework is established to process historical data. User behavior data is divided into different time scales such as minute level (instant operation), hour level (scene migration), week level (periodic rule), etc., and a feature association map across time dimensions is constructed, such as identifying an advertisement response difference model between the morning commuting period and the evening leisure period. The performance of user behavior data also varies at different time granularities. For example, minute-level data reflects instant operation behaviors, hour-level data reflects scene migration characteristics, and week-level data reveals periodic rules. In order to comprehensively capture these characteristics, a multi-granularity time analysis framework needs to be constructed to divide user behavior data into different time scales and establish a feature association map across time dimensions. In this process, how to accurately identify the relevance of user behaviors at different time scales and dynamically adjust the parameter ratio in feature engineering is the core problem in technical implementation. In particular, how to identify the advertisement response difference between the morning commuting period and the evening leisure period through the model and optimize the pushing strategy accordingly is the key to the technical problem. Summary of the Invention

[0004] The present invention provides an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis, mainly including:

[0005] Obtain multi-dimensional behavioral data of users in scenarios such as information browsing and short video viewing, and extract original features such as search terms, browsing duration, likes and comments; for the extracted original features, construct behavioral feature views with different time granularities such as second-level, minute-level, and hour-level to form multi-time scale feature representations; in the feature view of each time granularity, calculate the time weight distribution of feature nodes, and use an exponential decay function to perform time decay processing on historical features; analyze the feature views at different time scales through graph mining algorithms, mine the association rules between feature nodes, and construct a cross-scale feature association map; in the feature association map, according to the spectral clustering results of feature nodes and the gravitational direction of feature edges, judge the time correlation degree of different feature combinations; if the time correlation degree of a feature combination exceeds the empirical threshold, assign a high weight coefficient with a power-law distribution to the time weight of this feature combination; according to the vertical dimension and semantic level of features, dynamically adjust the network depth and connection density of the association map, and optimize the feature weight distribution in the map; based on the optimized feature weight distribution scheme, select the multi-path feature combination with the highest weight from the feature association map to generate recommended content for interest segmentation; match the generated personalized recommended content with the user's real-time feedback behavior, adjust the recommendation strategy, and continuously improve the conversion rate of the recommendation.

[0006] The present invention provides an artificial intelligence advertising push system based on multi-dimensional historical data analysis, mainly including: a data collection module for obtaining multi-dimensional behavior data of users in scenarios such as information browsing and short video viewing, and extracting original features such as search terms, browsing duration, likes and comments; a feature extraction module for constructing behavior feature views with different time granularities such as second-level, minute-level, and hour-level for the extracted original features to form multi-time scale feature representations; a time granularity view construction module for calculating the time weight distribution of feature nodes in the feature view of each time granularity and performing time decay processing on historical features using an exponential decay function; a time weight calculation module for analyzing the feature views at different time scales through a graph mining algorithm, mining the association rules between feature nodes, and constructing a cross-scale feature association graph; a graph mining analysis module for judging the time correlation degree of different feature combinations according to the spectral clustering results of feature nodes and the gravitational direction of feature edges in the feature association graph; a feature association graph construction module for assigning a high weight coefficient with a power-law distribution to the time weight of the feature combination if the time correlation degree of the feature combination exceeds an empirical threshold; a time correlation judgment module for dynamically adjusting the network depth and connection density of the association graph according to the vertical dimension and semantic level of the features and optimizing the feature weight distribution in the graph; a weight optimization module for selecting the multi-path feature combination with the highest weight from the feature association graph based on the optimized feature weight distribution scheme to generate interest-segmented recommended content; a recommended content generation module for matching the generated personalized recommended content with the user's real-time feedback behavior, adjusting the recommendation strategy, and continuously improving the conversion rate of the recommendation.

[0007] The technical solution provided by the embodiment of the present invention may include the following beneficial effects: The present invention discloses a personalized content recommendation method based on multi-dimensional behavior data. The method obtains the behavior data of users in scenarios such as information browsing and short video viewing, extracts feature views with multiple time granularities, and constructs a cross-scale feature association graph using a graph mining algorithm. The present invention judges the time correlation of feature combinations according to the spectral clustering results of feature nodes and the gravitational direction of feature edges, and dynamically adjusts the network depth and connection density of the graph. By selecting the multi-path feature combination with the highest weight, interest-segmented recommended content is generated, and the recommendation strategy is continuously optimized in combination with the user's real-time feedback behavior. The present invention realizes the accurate characterization of user interests and the intelligent recommendation of personalized content, effectively improving the conversion rate of the recommendation system and the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a flowchart of an artificial intelligence advertising push method based on multi-dimensional historical data analysis of the present invention.

[0009] Figure 2Schematic diagram of an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis according to the present invention.

[0010] Figure 3 Another schematic diagram of an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis according to the present invention.

[0011] Figure 4 Schematic structural diagram of an artificial intelligence advertisement pushing system based on multi-dimensional historical data analysis according to the present invention. Detailed implementation manners

[0012] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0013] As Figures 1 - 3 , an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis in this embodiment may specifically include:

[0014] Step S101, obtain multi-dimensional behavior data of users in scenarios such as information browsing and short video watching, and extract original features such as search terms, browsing duration, likes and comments.

[0015] Obtain the behavior data of at least one user in the scenarios of information browsing and short video watching. The behavior data includes search terms, browsing duration, number of likes, number of comments, viewing frequency, stay time, click-through rate, number of shares, number of collections, number of follows, interaction rate, and interest points; input the behavior data into a preset feature engineering model, and use the principal component analysis method to reduce the dimension of the behavior data. If there is a strong correlation between the features in the behavior data, they are combined into composite features. If the contribution degree of the features in the behavior data to the target is lower than a preset threshold, the feature is removed.

[0016] Specifically, the user's behavioral data in the information browsing and short video watching scenarios are obtained, and the behavioral data include search terms, browsing time, number of likes, number of comments, viewing frequency, dwell time, click-through rate, number of shares, number of favorites, number of follows, interaction rate and points of interest; for the behavioral data, data cleaning technology is used to remove noise data, if the data is missing, it is supplemented by interpolation method, if the data is abnormal, it is filtered through a preset threshold; the cleaned behavioral data is input into a preset feature engineering model, and the principal component analysis method is used to reduce the dimension of the behavioral data. If there is a strong correlation between the features in the behavioral data, they are merged into composite features. If the features in the behavioral data have a low contribution to the target, they are eliminated; based on the reduced-dimensional behavioral data, a user behavior profile is constructed, and if the user's browsing time exceeds a preset threshold, the user is marked For deep browsing users, if the number of likes and comments of the user are higher than the average, the user is marked as a high-interaction user; the user behavior portrait is input into the preset classification algorithm, and the random forest model is used to classify the users. If the user is a deep browsing user and a high-interaction user, the user is marked as a high-value user; if the user is a low-stay time and low-interaction user, the user is marked as a potential churn user; according to the classification results, a collaborative filtering algorithm is used to generate a personalized recommendation list. If the user is a high-value user, high-heat content is recommended first; if the user is a potential churn user, diversified content is recommended to enhance user interest; the personalized recommendation list is pushed to the user terminal, and the user portrait is updated using a real-time feedback mechanism. If the user clicks on the recommended content, the user's interest points are updated; if the user ignores the recommended content, the recommendation strategy is adjusted.

[0017] Step S102, constructing behavioral feature views at different time granularities such as second level, minute level, hour level, etc. for the extracted original features to form a multi-time scale feature representation.

[0018] Obtain the user's behavioral data in the scenarios of information browsing and short video watching, and build behavioral feature views at multiple time granularities such as seconds, minutes, and hours for the search terms, browsing time, number of likes, number of comments, viewing frequency, dwell time, click-through rate, number of shares, number of favorites, number of follows, interaction rate, and points of interest in the behavioral data to form a multi-time scale feature representation.

[0019] Specifically, users will leave a wealth of behavioral trajectory data in online scenarios. Constructing feature views with multiple time granularities can fully describe users' interests and intentions. For short video scenarios, the user's stay status per second, such as play, pause, fast forward, etc., can be counted to analyze the user's immediate reaction to the content. Taking entertainment short videos as an example, when watching celebrity dance videos, if users repeatedly watch a certain action clip many times, it means that these clips are more attractive to users. At the minute level, users' interactive behaviors such as likes, comments, and shares can be aggregated. For example, for food videos, users finish watching and like and share within two minutes, indicating that the content has resonated strongly with users. For educational videos, users may need longer to digest the content. At this time, attention should be paid to learning-related behavior features such as multiple pauses and replays. Hourly features reflect users' deep-seated interests. Taking information browsing as an example, users browse an average of ten articles per hour in the financial section, and stay for a long time, indicating that users are more concerned about financial investment. By analyzing users' search terms at different times, such as browsing financial news in the morning and paying attention to stock market comments in the evening, users' reading habits and interest changes can be identified. Interaction rate and click-through rate need to be calculated at different time scales. For news information, the ratio of user clicks to impressions per hour can be counted to reflect user activity. For short videos, the frequency of comment interactions at the minute level can be calculated to reflect the interactive value of the content. Collection and attention behaviors are suitable for analysis over a longer period of time because these behaviors indicate the user's continued interest. The core of constructing multi-time scale features is to describe the temporal evolution of user behavior. Second-level features can quickly perceive the user's current interest, minute-level features can reflect the user's attitude towards the content, and hour-level features can help grasp the user's long-term interest trend. This multi-dimensional feature representation provides a rich information basis for subsequent personalized recommendations and user analysis. Taking the live broadcast of sports events as an example, the second-level playback status can reflect the user's attention to the highlights, the minute-level barrage interaction shows the audience's emotional changes, and the hour-level viewing time shows the user's involvement in the game. These features of different time granularities complement each other and jointly construct a panoramic view of user behavior. For information with timeliness, the user's behavioral characteristics often change rapidly over time, so it is necessary to update the feature view in real time to capture the evolution of user interests in a timely manner.

[0020] Step S103, in the feature view of each time granularity, the time weight distribution of the feature nodes is calculated, and the exponential decay function is used to perform time decay processing on the historical features.

[0021] Construct behavioral feature views with multiple time granularities such as second-level, minute-level, and hour-level, and use the exponential decay function to calculate the time weight distribution of each time granularity feature node. According to the preset decay coefficient, perform time decay processing on the historical features of the second-level, minute-level, and hour-level feature nodes to obtain the decayed feature values. For the decayed second-level, minute-level, and hour-level feature values, use the weighted average algorithm to calculate the comprehensive score of each feature.

[0022] Specifically, the core idea of constructing the multi-time granularity behavior feature view lies in comprehensively depicting the user's behavior performance in different time dimensions. In the game application scenario, the player's operation behavior reflects the game experience at different time granularities. The second-level feature nodes record the player's immediate operations, such as skill releases and item uses. The minute-level feature nodes count the game behaviors with a longer time span, such as the number of combat rounds and the number of levels passed. The hour-level feature nodes reflect the player's game habits and persistence, including the daily login duration and the number of times of participating in activities. In the calculation of the time weight distribution, the exponential decay function can effectively balance the influence of historical behaviors. Taking the equipment purchase behavior of game players as an example, assuming the decay coefficient is 0.5, the weight of the purchase behavior one hour ago is 0.5, and that two hours ago is 0.25, and so on. This way of weight allocation ensures that the most recent behaviors have higher reference value, while also retaining the influence of historical behaviors. For the social media scenario, the user's interaction data also requires feature representations at multiple time scales. The second-level features capture the user's immediate reactions, such as the sliding speed and the stay duration. If the user stays on a certain piece of information for three consecutive seconds, it indicates that the content is highly attractive. The minute-level features record social interactions such as likes and comments. For example, if a user completes three forwards within five minutes, it shows that the content has good dissemination. The hour-level features reflect the user's activity patterns, such as the interaction frequency from 8 pm to 10 pm being significantly higher than other periods. Feature decay processing can highlight the importance of recent behaviors. In the news recommendation scenario, the weight of the user's reading preference in the most recent hour may be twice that six hours ago, reflecting the rapid decay of time-sensitive content. Through decay processing, the system can more accurately grasp the user's interest changes. For example, in a live sports event, the user's attention to the game result will rapidly decrease over time. The weighted average algorithm reasonably integrates the features at different time granularities. In the e-commerce scenario, the user's shopping intention is often reflected by multiple time dimensions together. The second-level stay time on the product detail page, the minute-level add-to-cart behavior, and the hour-level search history. The comprehensive score obtained after calculating the weights of these features can more comprehensively predict the user's purchase tendency. By reasonably setting the weights, both the real-time nature is ensured, and the stable expression of the user's long-term interests is maintained. This multi-dimensional time series feature representation provides a more reliable decision-making basis for personalized recommendations. In educational applications, the time features of learning behaviors have unique decay laws. The second-level features reflect the learning concentration, such as the pause frequency when watching teaching videos. The minute-level features show the knowledge absorption process, such as the speed of completing exercises. The hour-level features indicate the learning persistence, such as the learning input at a fixed time every day.

[0023] Step S104, analyze the feature views at different time scales through the graph mining algorithm, mine the association rules between feature nodes, and construct a cross-scale feature association graph.

[0024] Use the graph mining algorithm to analyze the second-level, minute-level, and hour-level feature views, and obtain the association rules between feature nodes. Through the mined association rules, construct a cross-scale feature association graph to describe the association relationship between feature nodes. If there are strongly associated nodes in the feature association graph, mark them as key feature nodes. According to the weight distribution of the key feature nodes, adjust the time granularity priority in the feature view. Through the adjusted feature view, optimize the parameters of the graph mining algorithm to improve the accuracy of association rule mining.

[0025] Specifically, the graph mining algorithm plays a key role in multi-time granularity feature analysis, and constructs a feature association graph by exploring the association patterns between nodes. In social applications, there is often a significant correlation between the second-level stay duration of users and the minute-level interaction behaviors. When a user continuously stays on a piece of content for more than ten seconds, the probability of generating comments or forwards within the next three minutes will increase by 60%. This association rule indicates the close connection between content attractiveness and user interaction willingness. For e-commerce platforms, the cross-scale feature association graph can reveal the user's shopping decision-making chain. The second-level browsing behavior of users on the product detail page and the hour-level purchase conversion rate form strongly associated nodes. Data shows that when the user's viewing duration of the product picture exceeds twenty seconds and the number of return views is greater than three times, the purchase probability within two hours will increase to 40%. Such high-weight associated nodes are marked as key features to guide the decision-making of the recommendation system. In the education scenario, the cross-time scale association of learning behaviors is particularly obvious. The second-level pause behavior of students when watching teaching videos is closely related to the hour-level knowledge mastery. When the number of pauses of learners significantly increases in difficult chapters, the system will increase the weight priority of the features in this time period, so as to more accurately evaluate the learning effect. In the news and information field, the reading path of users often shows clear temporal associations. The second-level reading speed change and the minute-level topic interest transfer form associated nodes. When the reading speed of users in financial news slows down by 30% and the stay time exceeds the normal level, the system will increase the weight of the feature nodes related to finance and optimize the subsequent content distribution. By continuously adjusting the time granularity priority of the feature view, the graph mining algorithm can more accurately capture the user behavior pattern. In music applications, the listening habits of users show unique temporal feature associations. When it is detected that the user has the second-level behavior of continuously skipping songs during the morning commute period, the system will enhance the feature weight of this period and optimize the playlist recommendation strategy accordingly. This dynamic adjustment mechanism based on association rules can improve the accuracy of the user experience.

[0026] Step S105, in the feature association graph, according to the spectral clustering results of the feature nodes and the gravitational direction of the feature edges, judge the time correlation degree of different feature combinations.

[0027] For the node set in the feature association graph, the spectral clustering algorithm is used to cluster and group the feature nodes to obtain the grouping values. According to the gravitational direction of the feature edges, the combination point relationship between different feature combinations is calculated to obtain the gravitational intensity of the combination points. Combining the grouping values and the combination point relationship, the time-degree distribution between different feature groups is judged to generate the time correlation degree. If the time correlation degree is higher than the preset threshold, mark this feature group as a highly correlated combination and extract the feature nodes in the highly correlated combination. Based on the feature nodes of the highly correlated combination, the time granularity priority in the feature view is adjusted to generate an optimized feature view. Using the optimized feature view, recalculate the gravitational direction of the feature edges and update the gravitational intensity in the feature association graph. Through the updated feature association graph, optimize the parameters of the spectral clustering algorithm to improve the accuracy of feature node clustering.

[0028] Specifically, the spectral clustering algorithm performs clustering and grouping through the similarity matrix of feature nodes, classifying nodes with similar behavioral characteristics. For example, in the e-commerce scenario, the similarity matrix of user browsing behavior nodes can be constructed based on dimensions such as dwell time and click depth to form user groups with different shopping intention strengths. When the average dwell time of a user on the product detail page exceeds three minutes and there are more than five price comparison behaviors, such nodes will be clustered into the high purchase intention group. The gravitational direction of the feature edges reflects the association strength between different feature nodes. In the news recommendation system, there is a significant gravitational relationship between the reading time of users for financial news and the stock trading period. When the stock market is open, the average dwell time of users on financial news increases by 50%, and the click-through rate of subsequent related articles increases to 30%, indicating a strong gravitational direction between these two feature nodes. The calculation of the combination point relationship is based on the interaction pattern between feature groups. In the video platform, the user's like behavior and comment interaction often form a close combination relationship. Data shows that when a user completes a like within one minute, the probability of leaving a comment within the next five minutes reaches 40%, and such a highly correlated combination point is given a greater gravitational intensity. The determination of the time correlation degree needs to consider the distribution law of feature groups at different time scales. The feature nodes of the highly correlated combination have an important impact on the time granularity priority of the feature view. After the feature view is optimized, the system needs to re-evaluate the gravitational direction of the feature edges. In the social platform, the content creation and interaction behaviors of users form a dynamic gravitational relationship network. When the golden interaction period after a creator publishes new content is identified as 30 minutes, the system will accordingly adjust the feature weights of this time window to optimize the content distribution strategy. By continuously iteratively optimizing the parameters of the spectral clustering algorithm, the accuracy of feature node clustering is improved. In the travel scenario, the matching degree between the user's travel plan and the actual travel time reflects the clustering effect. When the system finds that the accuracy of travel plans during the morning rush hour has increased to 85%, the clustering parameters are further adjusted to make the grouping of feature nodes more accurate.

[0029] Step S106: If the temporal correlation degree of the feature combination exceeds the empirical threshold, assign the temporal weight of this feature combination to the high weight coefficient of the power-law distribution.

[0030] Use the spectral clustering algorithm to group the node set and obtain the clustering results of the feature nodes. Calculate the relationship between the combination points of different feature combinations according to the gravitational direction of the feature edges to obtain the gravitational strength. Combine the clustering results and the gravitational strength to judge the temporal correlation degree of the feature groups and generate the temporal degree distribution. If the temporal correlation degree exceeds the preset value, use the power-law distribution algorithm to assign a high weight coefficient to the temporal weight. According to the high weight coefficient, adjust the temporal granularity priority in the feature view to generate an optimized feature view. Use the optimized feature view to recalculate the gravitational direction of the feature edges and update the gravitational strength. Optimize the spectral clustering algorithm parameters through the updated gravitational strength to improve the clustering accuracy of the feature nodes.

[0031] Step S107: Dynamically adjust the network depth and connection density of the association graph according to the vertical dimension and semantic level of the features, and optimize the feature weight distribution in the graph.

[0032] Construct an initial association graph according to the vertical dimension and semantic level of the features. Use the hierarchical clustering algorithm to layer the graph nodes to obtain the node distribution at different levels. Calculate the network depth and connection density of the graph according to the node level distribution. If the network depth exceeds the preset threshold, adjust the hierarchical structure of the graph and optimize the connection density. Reassign the feature weights according to the optimized network depth and connection density to generate a weight distribution table. Use the dynamic programming algorithm combined with the weight distribution table to update the feature weights in the association graph. Adjust the node connection relationship of the graph through the updated feature weights to generate an optimized association graph.

[0033] Specifically, in feature correlation analysis, the vertical dimension can reflect the hierarchical characteristics of business attributes. For example, in the product classification system of an e-commerce platform, a multi-level semantic relationship is formed from the top-level product categories to specific product attributes. When constructing the initial correlation graph, features such as product descriptions, price ranges, and sales channels are used as nodes, and connection relationships are established through semantic similarity. The hierarchical clustering algorithm performs stratification by calculating the distances between nodes. For example, the price range in product features can be clustered according to high, medium, and low levels, and sales channels can be grouped according to online and offline. In clothing products, features such as fabric, style, and size form node distributions at different levels, and the network depth and connection density are obtained by calculating the correlation degree between nodes. When the network depth is too large, it may lead to a decrease in the efficiency of feature correlation analysis. Taking home appliance products as an example, if they are expanded according to multiple dimensions such as brand, category, function, and parameters, the formed network levels may exceed the preset five-layer threshold. At this time, some similar function features need to be merged. For example, the refrigeration power and power consumption are merged into an energy efficiency index to optimize the network structure. Weight allocation should consider the importance of features. For example, in mobile phone products, the weight of the price range and performance configuration should be higher than that of the appearance color. By establishing a weight allocation table, a weight of 0.8 can be assigned to core features such as processor performance, while the weight of the appearance color is set to 0.3. The dynamic programming algorithm can adaptively adjust the connection relationships of feature nodes according to the weight allocation table. In the optimized correlation graph, the connections of feature nodes are more reasonable. This weight-based connection relationship can more accurately reflect the correlation degree between features, which is helpful for subsequent product recommendations and marketing strategy formulation. Through this multi-level feature analysis method, the correlation relationships between product features can be better understood, providing data support for precision marketing and personalized recommendations. The optimized correlation graph can quickly locate key feature nodes, improving the efficiency and accuracy of feature analysis.

[0034] Step S108, based on the optimized feature weight allocation scheme, select the multi-path feature combination with the highest weight from the feature correlation graph to generate interest-segmented recommended content.

[0035] According to the optimized network depth and connection density, reallocate feature weights to generate a weight allocation table. Combine the dynamic programming algorithm with the weight allocation table to update the feature weights in the correlation graph. Through the updated feature weights, extract the multi-path feature combination with the highest weight. According to the extracted multi-path feature combination, generate interest-segmented recommended content.

[0036] Specifically, feature weight assignment is a key link in the optimization of the association graph, and the weights are redistributed to reflect the importance of different features. Market factors such as price range and brand positioning also need to be considered in the weight assignment table, and the brand weight of high-end brands should be increased to more than 0.7. The dynamic programming algorithm updates the feature weights iteratively. In the field of clothing products, the weights of features such as fabric texture, style design, and brand level will be adjusted according to the seasons. In summer, the weight of fabric breathability is increased to 0.8, while the warmth retention property is decreased to 0.3. The algorithm can automatically adjust the weights according to sales data. For example, if a certain dress becomes popular due to its unique cut, the weight of the design feature will be increased accordingly. The extraction of multi-path feature combinations needs to comprehensively consider the association strength between features. Taking beauty products as an example, efficacy ingredients, suitable skin types, and price positioning constitute a high-weight feature combination. The weight of hyaluronic acid and vitamins in the efficacy ingredients reaches 0.85 due to their high-frequency co-occurrence. The algorithm will identify that the combination path of efficacy ingredients plus suitable skin types has the highest weight and make subsequent recommendations accordingly. The personalized recommendation based on feature combinations should adapt to the changes in user interests. In home furnishings, decoration style, space size, and color matching form typical feature combinations. When the system discovers that the user has recently browsed Nordic-style furniture, it will extract high-weight feature combinations including simple design, log materials, and gray-white color matching. When the user clicks on a certain type of feature combination multiple times, the recommendation weight of this combination will be further increased. Recommendations in the food and beverage field pay more attention to the balance between taste and nutrition. After identifying the feature combination that the user prefers low sugar and high protein, the system will give priority to recommending healthy snacks that meet this feature. If the user often buys products of a certain brand with a specific taste, the weight of this feature combination will be increased accordingly. Through the dynamically updated weight assignment, the recommendation system can respond in a timely manner to the transfer of user interests and provide more accurate personalized content.

[0037] Step S109: Match the generated personalized recommendation content with the user's real-time feedback behavior, adjust the recommendation strategy, and continuously improve the conversion rate of the recommendation. The user's real-time feedback behavior specifically refers to the user's retention duration and sharing and forwarding behavior.

[0038] Obtain the user's retention duration and sharing and forwarding behavior data, and calculate the matching degree between the number of behaviors and the amount of features. Dynamically adjust the recommendation strategy value according to the matching degree, and update the weight value of the personalized recommendation content. Combine the weight value with the dynamic programming algorithm to recalculate the relationship between the recommended quantity and the conversion rate. If the conversion rate is lower than the preset threshold, then optimize the combination method of the amount of features and the strategy value. Generate new personalized recommendation content according to the optimized feature combination. Judge the matching degree between the recommended content and the user behavior through the real-time behavior data. If the matching degree improves, then determine this recommendation strategy as the final optimization plan.

[0039] Specifically, the user retention duration reflects the attractiveness of content and user stickiness, while sharing and forwarding reflect the value of social dissemination. By tracking the time users spend on the platform and their interaction behaviors, the feature matching degree can be calculated. For example, in a fashion shopping platform, it is found that for products with an average browsing duration of over ten minutes by users, the matching degree of their style features with the user profile is usually above 0.8. For beauty products with high-frequency sharing and forwarding, their ingredient and efficacy features often highly match the users' skin care needs. Dynamically adjusting the recommendation strategy requires balancing recommendation accuracy and content diversity. If it is found that the average retention duration of users for a certain category decreases, the system will correspondingly reduce the recommendation weight of that category. If users frequently share content related to sports equipment to their Moments, then the recommendation proportion of professional sports equipment will be increased. By adjusting the weight value in real time, it is ensured that the recommended content always fits the changes in users' interests. The relationship between the recommended quantity and the conversion rate reflects the recommendation effect. Taking maternal and child products as an example, when the system recommends 20 products per day, the conversion rate is 5%. After optimization through the dynamic programming algorithm, the recommended quantity of products with high matching degree is increased to 30, and the recommended quantity of products with low matching degree is reduced to 10, and the conversion rate is increased to 8%. This shows that precise targeted recommendations can enhance users' purchase willingness. The combination method of feature quantities and strategy values determines the recommendation effect. In the field of household appliances, the weight ratios of features such as price range, function configuration, and brand level will affect users' decisions. When it is found that the conversion rate of high-price brand refrigerators is lower than expected, the system will adjust the feature combination, increase the weights of practical features such as energy-saving rating and usage experience, and improve the recommendation effect. Real-time behavior data reflects users' acceptance of the recommended content. In book recommendations, behavioral indicators such as the reading duration and the number of notes after users receive recommendations can directly reflect the matching degree. If it is found that the reading completion rate of users for popular science books increases and they often add bookmarks and take notes, it indicates that the recommendation strategy for this type of book has a high matching degree. The system will solidify this strategy into an optimization plan and continuously use and fine-tune it in subsequent recommendations. The recommendation of fresh food needs to consider seasonal and local characteristics. When in-season fruits receive a high retention duration and sharing volume, the system will increase the recommendation weight of similar categories. At the same time, according to the purchase frequency and evaluations of users, the combination method of features such as origin and price range will be dynamically adjusted to continuously optimize the recommendation effect. By continuously monitoring users' feedback on different feature combinations, the recommendation system can establish a more precise personalized service model.

[0040] Step S1010: Integrate the feature weight distributions at multiple time scales, judge the time cycle mode in which the user is currently located, and adjust the weight ratio. If the user is in the commuting peak period on a weekday, then focus on the search hot words and real-time click feedback in the last hour, and increase the weight ratio of these features. If the user is in the late-night leisure period on a weekend, then focus on the topic preferences and video type distributions within a week, and increase the weight ratio of these features.

[0041] Obtain the user's behavior data at different time scales, including search hot words, click feedback, topic preferences, and video type distributions. Use pattern recognition algorithms to determine the time cycle patterns of the user during the commuting peak period or late-night leisure period. According to the period judgment results, dynamically adjust the weight ratio of the feature distribution to increase the weight of the behavior data corresponding to the period. Use clustering algorithms to classify the user's behavior data and generate feature combinations for different time cycles. According to the feature combinations, recalculate the user's interest preferences at different time periods. Use collaborative filtering algorithms to combine the user's interest preferences and behavior data to generate personalized recommendation content. Dynamically update the weight ratio of the recommendation content through the user's real-time behavior data to optimize the recommendation strategy.

[0042] Specifically, the collection of user behavior data on a time scale is crucial for recommendation systems. During the commuting period, urban users are more inclined to search for practical information such as breakfast shops and weather forecasts, and watch short videos with a viewing duration controlled within three to five minutes. While during the late-night leisure period, users will spend more time browsing entertainment content such as food and travel, and the average viewing duration can reach more than fifteen minutes. Pattern recognition algorithms can judge the time period by analyzing the active patterns of users. Taking a certain video platform as an example, during the period from 7:00 to 9:00 in the morning on weekdays, the search hot words of the commuting group mainly focus on fields such as news and traffic conditions. After 10:00 pm, the click-through rate of leisure and entertainment content increases significantly, and the acceptance of long videos by users also increases accordingly. The dynamic adjustment of feature distribution weights needs to consider multiple dimensions. During the morning rush hour, the weight of news content can be increased to 0.6, while the weights of relaxing content such as food and entertainment are reduced to 0.2. This adjustment can ensure that the pushed content matches the needs of users in the current time period. Behavioral data clustering analysis can reveal the time-period characteristics of users. A certain news and information platform found that office worker users tend to quickly browse the news summary in the morning on weekdays, are more willing to watch relaxing life topics during lunch breaks, and prefer in-depth reading and knowledge-based content on weekends. Based on these characteristics, the platform can push more highly matched content at different times. The calculation of user interest preferences needs to combine the time dimension. For example, during the morning commuting period, if a user frequently clicks on financial news, the system will increase the recommendation weight of such content to 0.7. When it is found that a user often watches food videos late at night, the recommendation weight of food-related content will be correspondingly increased to 0.5. Collaborative filtering algorithms play a key role in personalized recommendation. Data from an educational platform shows that the concentration of high school students watching teaching videos during self-study hours is very high, and the conversion rate of providing recommendations for the same type of courses can reach 30% at this time. On weekends, the same group has a higher acceptance of interest cultivation content, and personalized recommendations need to be adjusted in a timely manner. The dynamic update of real-time behavior data ensures that the recommended content always maintains freshness. A music platform found that there are obvious differences in users' music style preferences at different times. In the early morning, they tend to prefer light music, accounting for 40%, while at night, they prefer songs with a strong rhythm, accounting for up to 60%. By real-time adjusting the recommendation weights of music types, the user's stay duration can be significantly increased.

[0043] Such as Figure 4As shown in the figure, the embodiment of the present invention also provides an artificial intelligence advertising push system based on multi-dimensional historical data analysis, which mainly includes: a data collection module, used to obtain multi-dimensional behavior data of users in scenarios such as information browsing and short video viewing, and extract original features such as search terms, browsing duration, likes and comments; a feature extraction module, used to construct behavior feature views with different time granularities such as second-level, minute-level, and hour-level for the extracted original features, and form multi-time scale feature representations; a time granularity view construction module, used to calculate the time weight distribution of feature nodes in the feature view of each time granularity, and perform time decay processing on historical features using an exponential decay function; a time weight calculation module, used to analyze the feature views at different time scales through a graph mining algorithm, mine the association rules between feature nodes, and construct a cross-scale feature association graph; a graph mining analysis module, used to judge the time correlation degree of different feature combinations according to the spectral clustering results of feature nodes and the gravitational direction of feature edges in the feature association graph; a feature association graph construction module, used to assign the time weight of the feature combination to the high weight coefficient of the power-law distribution if the time correlation degree of the feature combination exceeds the empirical threshold; a time correlation judgment module, used to dynamically adjust the network depth and connection density of the association graph according to the vertical dimension and semantic level of the features, and optimize the feature weight distribution in the graph; a weight optimization module, used to select the multi-path feature combination with the highest weight from the feature association graph based on the optimized feature weight distribution scheme, and generate interest-segmented recommendation content; a recommendation content generation module, used to match the generated personalized recommendation content with the user's real-time feedback behavior, adjust the recommendation strategy, and continuously improve the conversion rate of the recommendation; a real-time feedback adjustment module, used to fuse the feature weight distributions at multiple time scales, judge the time cycle mode in which the user is currently located, adjust the weight ratio. If the user is in the commuting peak period on a weekday, focus on their search hot words and real-time click feedback in the last hour, and increase the weight ratio of these features. If the user is in the late-night leisure period on the weekend, focus on their topic preferences and video type distributions within a week, and increase the weight ratio of these features. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. An artificial intelligence advertisement push method based on multi-dimensional historical data analysis, characterized in that, The method includes: Obtain multi-dimensional behavioral data of users in different scenarios and extract original feature data; for the extracted original features, construct behavioral feature views with different time granularities of seconds, minutes, and hours to form multi-time scale feature representations; in the feature views of each time granularity, calculate the time weight distribution of feature nodes, and use an exponential decay function to perform time decay processing on historical features; analyze the feature views at different time scales through graph mining algorithms, mine the association rules between feature nodes, and construct a cross-scale feature association graph; in the feature association graph, according to the spectral clustering results of feature nodes and the gravitational direction of feature edges, judge the time correlation degree of different feature combinations; if the time correlation degree of a feature combination exceeds the empirical threshold, assign the time weight of this feature combination to the high weight coefficient of the power-law distribution; according to the vertical dimension and semantic level of features, dynamically adjust the network depth and connection density of the association graph, and optimize the feature weight distribution in the graph; based on the optimized feature weight distribution scheme, select the multi-path feature combination with the highest weight from the feature association graph to generate recommended content for interest segmentation; match the generated personalized recommended content with the real-time feedback behavior of users, adjust the recommendation strategy, and continuously improve the conversion rate of recommendations.

2. The method according to claim 1, wherein The obtaining of multi-dimensional behavioral data of users in different scenarios and the extraction of original feature data include: Obtain the behavioral data of at least one user in the scenarios of information browsing and short video viewing, where the behavioral data includes search terms, browsing duration, number of likes, number of comments, viewing frequency, stay time, click-through rate, number of shares, number of collections, number of follows, interaction rate, and points of interest. Input the behavioral data into a preset feature engineering model, and use the principal component analysis method to reduce the dimension of the behavioral data. If there is a strong correlation between the features in the behavioral data, they are combined into composite features. If the contribution degree of a feature in the behavioral data to the target is lower than the preset threshold, the feature is removed.

3. The method according to claim 1, wherein The constructing of behavioral feature views with different time granularities of seconds, minutes, and hours for the extracted original features to form multi-time scale feature representations includes: Obtain the behavioral data of users in the scenarios of information browsing and short video viewing, and construct behavioral feature views with multi-time granularities of seconds, minutes, and hours for the search terms, browsing duration, number of likes, number of comments, viewing frequency, stay time, click-through rate, number of shares, number of collections, number of follows, interaction rate, and points of interest in the behavioral data to form multi-time scale feature representations.

4. The method according to claim 3, wherein The calculating of the time weight distribution of feature nodes in the feature view of each time granularity and the performing of time decay processing on historical features using an exponential decay function include: Construct behavioral feature views with multi-time granularities of seconds, minutes, and hours, and use an exponential decay function to calculate the time weight distribution of feature nodes in each time granularity. According to the preset decay coefficient, perform time decay processing on the historical features of feature nodes in seconds, minutes, and hours to obtain the decayed feature values. For the decayed second-level, minute-level, and hour-level eigenvalue, the weighted average algorithm is adopted to calculate the comprehensive score of each feature.

5. The method according to claim 4, wherein The analysis of the feature views at different time scales by the graph mining algorithm, mining the association rules between feature nodes, and constructing a cross-scale feature association map includes: Adopt the graph mining algorithm to analyze the second-level, minute-level, and hour-level feature views, and obtain the association rules between feature nodes; Through the mined association rules, construct a cross-scale feature association map to describe the association relationship between feature nodes; If there are strongly associated nodes in the feature association map, mark them as key feature nodes; According to the weight distribution of the key feature nodes, adjust the time granularity priority in the feature view; Through the adjusted feature view, optimize the parameters of the graph mining algorithm to improve the accuracy of association rule mining.

6. The method according to claim 1, characterized in that, In the feature association map, according to the spectral clustering results of the feature nodes and the gravitational direction of the feature edges, judge the time correlation degree of different feature combinations, including: For the node set in the feature association map, use the spectral clustering algorithm to cluster and group the feature nodes to obtain the grouping values; According to the gravitational direction of the feature edges, calculate the combination point relationship between different feature combinations to obtain the gravitational intensity of the combination points; Combined with the grouping values and the combination point relationship, judge the time degree distribution between different feature groups to generate the time correlation degree; If the time correlation degree is higher than the preset threshold, mark this feature group as a highly correlated combination and extract the feature nodes in the highly correlated combination; According to the feature nodes of the highly correlated combination, adjust the time granularity priority in the feature view to generate an optimized feature view; Use the optimized feature view to recalculate the gravitational direction of the feature edges and update the gravitational intensity in the feature association map; Through the updated feature association map, optimize the parameters of the spectral clustering algorithm to improve the accuracy of feature node clustering.

7. The method according to claim 6, characterized in that, If the time correlation degree of the feature combination exceeds the empirical threshold, assign the time weight of this feature combination to the high weight coefficient of the power-law distribution, including: Use the spectral clustering algorithm to group the node set to obtain the clustering results of the feature nodes; According to the gravitational direction of the feature edges, calculate the relationship between the combination points of different feature combinations to obtain the gravitational intensity; Combined with the clustering results and the gravitational intensity, judge the time correlation degree of the feature group to generate the time degree distribution; If the time correlation degree exceeds the preset value, use the power-law distribution algorithm to assign a high weight coefficient to the time weight; According to the high weight coefficient, adjust the time granularity priority in the feature view to generate an optimized feature view.

8. The method according to claim 1, characterized in that According to the vertical dimension and semantic level of the feature, dynamically adjust the network depth and connection density of the association map, and optimize the feature weight distribution in the map, including: Construct an initial association map according to the vertical dimension and semantic level of the feature; Use the hierarchical clustering algorithm to stratify the map nodes to obtain the node distribution of different levels; According to the node level distribution, calculate the network depth and connection density of the map; If the network depth exceeds the preset threshold, adjust the hierarchical structure of the map and optimize the connection density; According to the optimized network depth and connection density, reassign the feature weights to generate a weight distribution table; The dynamic programming algorithm is combined with the weight distribution table to update the feature weights in the association graph; Based on the updated feature weights, adjust the node connection relationship of the graph to generate an optimized association graph.

9. The method according to claim 1, characterized in that, Based on the optimized feature weight distribution scheme, select the multi-path feature combination with the highest weight from the feature association graph to generate interest-segmented recommended content, including: According to the optimized network depth and connection density, re-distribute the feature weights to generate a weight distribution table; The dynamic programming algorithm is combined with the weight distribution table to update the feature weights in the association graph; Based on the updated feature weights, extract the multi-path feature combination with the highest weight; According to the extracted multi-path feature combination, generate interest-segmented recommended content.

10. The method according to claim 1, wherein Match the generated personalized recommended content with the user's real-time feedback behavior, adjust the recommendation strategy, and continuously improve the conversion rate of the recommendation. The user's real-time feedback behavior specifically refers to the user's retention duration and sharing and forwarding behavior, including: Obtain the user's retention duration and sharing and forwarding behavior data, and calculate the matching degree between the behavior number and the feature quantity; Dynamically adjust the recommendation strategy value according to the matching degree, and update the weight value of the personalized recommended content; Use the dynamic programming algorithm combined with the weight value to recalculate the relationship between the recommended quantity and the conversion rate; If the conversion rate is lower than the preset threshold, optimize the combination method of the feature quantity and the strategy value; Generate new personalized recommended content according to the optimized feature combination; Judge the matching degree between the recommended content and the user behavior through real-time behavior data; If the matching degree is improved, determine this recommendation strategy as the final optimization plan.

Images