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

Abstract

The application provides an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis, comprising: constructing behavior feature views of different time granularities such as seconds, minutes and hours for the extracted original features to form multi-time scale feature representation; analyzing the feature views under different time scales through a graph mining algorithm, mining the association rules between feature nodes, and constructing a cross-scale feature association graph; in the feature association graph, judging the time correlation degree of different feature combinations according to the spectral clustering results of the feature nodes and the gravitational direction of the feature edges; if the time correlation degree of the feature combination exceeds an empirical threshold, assigning the time weight of the feature combination to a high weight coefficient of a power law distribution; 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.

Description

Technical Field

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

[0002] In ad push decisions, effectively quantifying the contribution of data from different time dimensions is a complex technical problem. User behavior data has multi-layered temporal characteristics, such as long-term browsing preferences, short-term search history, and seasonal cyclical behavior. The influence of this data on ad push varies significantly across different time scales. To accurately capture these differences, it is necessary to construct a weight calculation model based on an attention mechanism that dynamically evaluates the contribution of each time dimension to the current push. However, due to the time decay effect of data, the impact of historical data from different time dimensions on the current decision is not linear, but rather gradually weakens over time. Therefore, it is necessary to introduce a time decay factor to more accurately reflect the timeliness of historical data.

[0003] Furthermore, ad push decisions require quantifying the contribution of data from different historical dimensions. For example, this involves analyzing data layers such as users' long-term browsing preferences, short-term search history, and seasonal cyclical behavior. A contribution evaluation model based on a time decay factor can be established to dynamically adjust the parameter ratios in feature engineering. A multi-granularity temporal analysis framework is also needed to process historical data. User behavior data can be segmented into different time scales, such as minutes (for immediate actions), hours (for scenario transitions), and weeks (for cyclical patterns). A feature correlation map across time dimensions can be constructed. For example, a model can be used to identify differences in ad response between morning commutes and evening leisure time. User behavior data also exhibits differences at different time granularities. For example, minute-level data reflects immediate actions, hour-level data reflects scenario transitions, and weekly data reveals cyclical patterns. To fully capture these characteristics, a multi-granularity temporal analysis framework is needed. This framework segments user behavior data into different time scales and constructs a feature correlation map across time dimensions. Accurately identifying the correlations between user behavior at different time scales and dynamically adjusting the parameter ratios in feature engineering are key technical challenges. In particular, the key technical issue is how to use models to identify the differences in advertising responses between the morning commuting period and the evening leisure period, and optimize the push strategy accordingly. Summary of the Invention

[0004] The present invention provides an artificial intelligence advertising push method based on multi-dimensional historical data analysis, which mainly includes:

[0005] Obtain multi-dimensional behavioral data of users in scenarios such as information browsing and short video watching, and extract original features such as search terms, browsing time, likes and comments; Based on the extracted original features, construct behavioral feature views at different time granularities such as seconds, minutes, and hours to form a multi-time scale feature representation; 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 feature views at different time scales through graph mining algorithms, mine association rules between feature nodes, and construct a cross-scale feature association graph; In the feature association graph, determine the temporal correlation of different feature combinations based on the spectral clustering results of feature nodes and the gravitational direction of feature edges; If the temporal correlation of a feature combination exceeds an empirical threshold, assign the time weight of the feature combination a high weight coefficient of the power law distribution; Dynamically adjust the network depth and connection density of the association graph based on the vertical dimension and semantic level of the feature to 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 interest-based recommended content; 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 recommendations.

[0006] The application provides an artificial intelligence advertisement pushing system based on multi-dimensional historical data analysis, mainly comprising: a data acquisition module for acquiring multi-dimensional behavior data of a user in scenarios such as information browsing and short video watching, and extracting original features such as search words, browsing time length and likes and comments; a feature extraction module for constructing behavior feature views of different time granularities such as seconds, minutes and hours for the extracted original features, and forming multi-time scale feature representations; a time granularity view construction module for calculating the time weight distribution of feature nodes in each time granularity feature view, and performing time attenuation processing on historical features by using an exponential decay function; a time weight calculation module for analyzing the feature views under different time scales by using 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 in the feature association graph according to the spectral clustering results of the feature nodes and the gravitational direction of the feature edges; a feature association graph construction module for assigning the time weight of the feature combination to a high weight coefficient of a power law distribution 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, and generating recommended content of interest subdivision; and a recommended content generation module for matching the generated personalized recommended content with the real-time feedback behavior of the user, adjusting the recommendation strategy, and continuously improving the conversion rate of the recommendation.

[0007] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0008] The application discloses a personalized content recommendation method based on multi-dimensional behavior data. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 The application provides a flowchart of an artificial intelligence advertisement pushing method based on multi-dimensional historical data analysis.

[0010] Figure 2 This is a schematic diagram of an artificial intelligence advertising push method based on multi-dimensional historical data analysis of the present invention.

[0011] Figure 3 This is another schematic diagram of an artificial intelligence advertising push method based on multi-dimensional historical data analysis of the present invention.

[0012] Figure 4 This is a structural diagram of an artificial intelligence advertising push system based on multi-dimensional historical data analysis of the present invention. DETAILED DESCRIPTION

[0013] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0014] like Figure 1-3 This embodiment provides an artificial intelligence advertising push method based on multi-dimensional historical data analysis, which may specifically include:

[0015] Step S101: Obtain multi-dimensional behavioral data of users in scenarios such as information browsing and short video watching, and extract original features such as search terms, browsing time, likes and comments.

[0016] Obtain behavioral data of at least one user in information browsing and short video watching scenarios, wherein the behavioral data includes 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; 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 merged into a composite feature; if the contribution of a feature in the behavioral data to the target is lower than a preset threshold, the feature is eliminated.

[0017] 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 the 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 behavioral data after dimensionality reduction, a user behavior profile is constructed, and if the user's browsing time exceeds the 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; based on the classification results, a collaborative filtering algorithm is used to generate a personalized recommendation list. If the user is a high-value user, high-popularity 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 points of interest are updated; if the user ignores the recommended content, the recommendation strategy is adjusted.

[0018] In step S102 , based on the extracted original features, behavioral feature views at different time granularities, such as second level, minute level, hour level, etc., are constructed to form a multi-time scale feature representation.

[0019] Obtain user behavioral data in information browsing and short video viewing scenarios, and construct behavioral feature views at multiple time granularities such as seconds, minutes, and hours based on 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.

[0020] Specifically, users leave behind a wealth of behavioral data in online scenarios. Constructing feature views with multiple time granularities can comprehensively describe user interests and intentions. For short videos, we can count users' pause states per second, such as play, pause, and fast-forward, to analyze their immediate reactions to the content. For example, when watching a celebrity dance video, if users repeatedly watch a certain action segment, it indicates that these segments are more engaging. At the minute level, we can aggregate user interactions such as likes, comments, and shares. For example, if users complete a food video and like and share it within two minutes, it indicates that the content resonates strongly with users. For educational videos, users may take longer to digest the content, so we should focus on learning-related behavior features such as frequent pauses and replays. Hourly features reflect users' deeper interests. For example, in the financial section, users browse an average of ten articles per hour and spend a long time on the content, indicating a strong interest in financial investment. By analyzing users' search terms at different times of day, such as browsing financial news in the morning and focusing on stock market commentary in the evening, we can identify changes in users' reading habits and interests. Interaction rates and click-through rates need to be calculated at different time scales. For news, the ratio of user clicks to impressions per hour can be calculated to reflect user activity. For short videos, the frequency of comment interactions can be calculated at the minute level to reflect the interactive value of the content. Favoriting and following behaviors are more suitable for analysis over longer time periods, as they indicate sustained user interest. The key to constructing multi-time-scale features is to describe the temporal evolution of user behavior. Second-level features can quickly perceive a user's current interests, minute-level features can reflect user attitudes toward content, and hour-level features can help identify long-term trends in user interests. This multi-dimensional feature representation provides a rich information foundation for subsequent personalized recommendations and user analysis. For live sports events, for example, second-level playback status can reflect user attention to highlights, minute-level comment interactions can reveal audience emotional shifts, and hour-level viewing duration can indicate user engagement with the game. These features at different time granularities complement each other, constructing a comprehensive view of user behavior. For time-sensitive information, user behavior characteristics often change rapidly over time, necessitating real-time feature updates to capture the evolving patterns of user interests.

[0021] Step S103 : In the feature view of each time granularity, the time weight distribution of the feature nodes is calculated, and an exponential decay function is used to perform time decay processing on the historical features.

[0022] We construct behavioral feature views at multiple time granularities, such as seconds, minutes, and hours. We use an exponential decay function to calculate the time weight distribution of feature nodes at each time granularity. Based on the preset decay coefficient, we perform time decay on the historical features of feature nodes at the second, minute, and hour levels to obtain the decayed feature values. We then use a weighted average algorithm to calculate the composite score for each feature using these decayed second, minute, and hourly feature values.

[0023] Specifically, the core concept behind constructing multi-time-granularity behavioral feature views is to comprehensively characterize user behavior across different time dimensions. In gaming scenarios, player actions reflect the gaming experience at different time scales. Second-level feature nodes record immediate player actions, such as skill activation and item usage. Minute-level feature nodes count longer-term gaming behaviors, such as combat rounds and level completion. Hour-level feature nodes reflect player gaming habits and persistence, including daily login time and number of activity participation. In calculating temporal weight distribution, an exponential decay function effectively balances the influence of historical behavior. Taking a player's equipment purchase behavior as an example, assuming a decay coefficient of 0.5, a purchase made one hour ago would have a weight of 0.5, a purchase made two hours ago would have a weight of 0.25, and so on. This weight distribution ensures that recent behavior has a higher reference value while also preserving the influence of historical behavior. In social media scenarios, user interaction data also requires multi-time-scale feature representation. Second-level features capture immediate user reactions, such as scrolling speed and dwell time. If a user dwells on a post for three consecutive seconds, the content is highly engaging. Minute-level features record social interactions such as likes and comments. For example, if a user reposts a post three times within five minutes, it indicates high virality. Hour-level features reflect user activity patterns, such as significantly higher interaction frequency between 8:00 PM and 10:00 PM than at other times. Feature decay processing can emphasize the importance of recent behavior. In news recommendations, a user's reading preference in the last hour may be twice as weighted as that from six hours ago, reflecting the rapid decay of time-sensitive content. This decay processing allows the system to more accurately grasp evolving user interests. For example, in live sports broadcasts, user interest in match results decreases rapidly over time. The weighted average algorithm rationally integrates features of different temporal granularities. In e-commerce scenarios, user purchasing intent is often reflected across multiple temporal dimensions. A weighted composite score derived from features such as time spent on a product detail page (seconds), add-to-cart behavior (minutes), and search history (hours) can more comprehensively predict a user's purchasing intent. By appropriately assigning weights, real-time performance is ensured while maintaining a stable expression of user interests over time. This multi-dimensional temporal feature representation provides a more reliable basis for personalized recommendations. In educational apps, the temporal characteristics of learning behavior exhibit unique decay patterns. Second-level features reflect learning focus, such as the frequency of pauses while watching an instructional video. Minute-level features reveal the process of knowledge absorption, such as the speed of completing exercises. Hour-level features indicate learning continuity, such as the amount of learning time invested at a fixed time each day.

[0024] Step S104: Analyze feature views at different time scales through a graph mining algorithm, mine association rules between feature nodes, and construct a cross-scale feature association graph.

[0025] A graph mining algorithm is used to analyze feature views at the second, minute, and hour levels to extract association rules between feature nodes. Based on the mined association rules, a cross-scale feature association graph is constructed to describe the associations between feature nodes. If there are strongly associated nodes in the feature association graph, they are marked as key feature nodes. Based on the weight distribution of key feature nodes, the time granularity priority in the feature view is adjusted. Using this adjusted feature view, graph mining algorithm parameters are optimized to improve the accuracy of association rule mining.

[0026] Specifically, graph mining algorithms play a key role in analyzing multi-time granularity features, constructing feature correlation graphs by exploring inter-node association patterns. In social applications, there is often a significant correlation between user dwell time in seconds and minute-level interaction. When a user spends more than ten seconds on a piece of content, the probability of commenting or forwarding it within the next three minutes increases by 60%. This correlation pattern demonstrates a close connection between content appeal and user engagement. For e-commerce platforms, cross-scale feature correlation graphs can reveal the user shopping decision chain. Second-level browsing behavior on product detail pages is strongly correlated with hour-level purchase conversion rates. Data shows that when a user views a product image for more than twenty seconds and returns to view it more than three times, their probability of purchasing within two hours increases to 40%. These high-weighted correlation nodes are labeled as key features and used to guide recommendation system decisions. In educational settings, the cross-timescale correlation of learning behavior is particularly pronounced. Students' pauses in instructional videos, measured in seconds, are closely correlated with their knowledge mastery at the hour-level. When learners pause significantly more often during difficult chapters, the system prioritizes the features for that time period, allowing for more accurate assessment of learning outcomes. In the news and information sector, users' reading paths often exhibit clear temporal correlations. Second-level changes in reading speed and minute-level shifts in topic interest form associated nodes. When a user's reading speed for financial news slows by 30% and their dwell time exceeds the average, the system increases the weight of financial-related feature nodes to optimize subsequent content distribution. By continuously adjusting the temporal granularity priority of feature views, graph mining algorithms can more accurately capture user behavior patterns. In music apps, users' listening habits exhibit unique temporal feature correlations. When a user is detected skipping songs continuously for seconds during their morning commute, the system increases the feature weight for that time period and optimizes the playlist recommendation strategy accordingly. This dynamic adjustment mechanism based on association rules can improve the accuracy of the user experience.

[0027] Step S105 : in the feature association graph, the temporal correlation of different feature combinations is determined based on the spectral clustering results of the feature nodes and the gravitational directions of the feature edges.

[0028] For the node set in the feature correlation graph, a spectral clustering algorithm is used to cluster and group the feature nodes to obtain a grouping value. According to the attractive direction of the feature edge, the combination point relationship between different feature combinations is calculated to obtain the attractive strength of the combination point. In combination with the grouping value and the combination point relationship, the time degree distribution between different feature groups is judged to generate a time correlation degree. If the time correlation degree is higher than a preset threshold, the feature group is marked as a high correlation combination, and the feature nodes in the high correlation combination are extracted. According to the feature nodes of the high correlation combination, the time granularity priority in the feature view is adjusted to generate an optimized feature view. The attractive direction of the feature edge is recalculated using the optimized feature view, and the attractive strength in the feature correlation graph is updated. Through the updated feature correlation graph, the spectral clustering algorithm parameters are optimized to improve the accuracy of feature node clustering.

[0029] Specifically, the spectral clustering algorithm clusters and groups through the similarity matrix of the feature nodes, and nodes with similar behavior characteristics are classified. For example, in the e-commerce scenario, user browsing behavior nodes can construct a similarity matrix according to the dimensions of dwell time, click depth, etc., forming user groups with different shopping intention strengths. When the average dwell time of users on the product detail page exceeds three minutes, and there are more than five price comparison behaviors, such nodes will be clustered into a high purchase intention group. The attractive direction of the feature edge reflects the correlation strength between different feature nodes. In the news recommendation system, the reading time of users on financial news shows a significant attractive relationship with the stock trading period. When the stock market opens, the average dwell time of users on financial news increases by 50%, and the subsequent related article click rate increases to 30%, indicating that there is a strong attractive direction between the two feature nodes. The calculation of the combination point relationship is based on the interaction mode between feature groups. In the video platform, the like behavior of users and the comment interaction often form a close combination relationship. Data shows that when a user likes within one minute, the probability of publishing a comment within the next five minutes reaches 40%, and this highly correlated combination point is given a larger attractive strength. The determination of the time correlation degree needs to consider the distribution law of the feature group at different time scales. The feature nodes of the high correlation combination have an important influence on the time granularity priority of the feature view. After the optimization of the feature view, the system needs to reevaluate the attractive direction of the feature edge. In the social platform, the content creation and interaction behavior of users form a dynamic attractive relationship network. When the golden interaction period after the creator publishes new content is identified as thirty minutes, the system will adjust the feature weight of this time window accordingly to optimize the content distribution strategy. By continuously iterating and optimizing the spectral clustering algorithm parameters, the accuracy of feature node clustering is improved. In the travel scenario, the matching degree of the user's travel plan and the actual travel time reflects the clustering effect. When the system finds that the accuracy rate of travel planning during the morning rush hour has increased to 85%, the clustering parameters are further adjusted to make the grouping of feature nodes more accurate.

[0030] Step S106, if the time correlation degree of the feature combination exceeds the empirical threshold, the time weight of the feature combination is assigned to the high weight coefficient of the power-law distribution.

[0031] The spectral clustering algorithm is used to group the node set, and the clustering result of the feature node is obtained. According to the attractive direction of the feature edge, the relationship between the combination points of different feature combinations is calculated, and the attractive strength is obtained. Combining the clustering result and the attractive strength, the time correlation degree of the feature group is judged, and the time degree distribution is generated. If the time correlation degree exceeds the preset value, the power-law distribution algorithm is used to assign a high weight coefficient to the time weight. According to the high weight coefficient, the time granularity priority in the feature view is adjusted, and the optimized feature view is generated. The optimized feature view is used to recalculate the attractive direction of the feature edge, and the attractive strength is updated. Through the updated attractive strength, the spectral clustering algorithm parameters are optimized, and the clustering accuracy of the feature node is improved.

[0032] Step S107, according to the vertical dimension and semantic hierarchy of the feature, the network depth and connection density of the association graph are dynamically adjusted, and the feature weight distribution in the graph is optimized.

[0033] According to the vertical dimension and semantic hierarchy of the feature, an initial association graph is constructed. The hierarchical clustering algorithm is used to layer the graph nodes, and the node distribution of different levels is obtained. According to the node level distribution, the network depth and connection density of the graph are calculated. If the network depth exceeds the preset threshold, the level structure of the graph is adjusted, and the connection density is optimized. According to the optimized network depth and connection density, the feature weight is redistributed, and the weight distribution table is generated. The dynamic programming algorithm is used in combination with the weight distribution table to update the feature weight in the association graph. Through the updated feature weight, the node connection relationship of the graph is adjusted, and the optimized association graph is generated.

[0034] Specifically, in feature association analysis, vertical dimensions can reflect the hierarchical nature of business attributes. For example, in the product classification system of e-commerce platforms, multi-level semantic relationships are formed from the top-level product categories to specific product attributes. When constructing the initial association graph, features such as product description, price range, and sales channel are used as nodes, and connections are established based on semantic similarity. Hierarchical clustering algorithms calculate the distance between nodes to achieve stratification. For example, within product features, price ranges can be clustered into high, medium, and low-end categories, and sales channels can be grouped into online and offline categories. In clothing products, features such as fabric, style, and size form nodes at different levels. The correlation between nodes is calculated to determine network depth and connection density. Excessive network depth can reduce the efficiency of feature association analysis. For example, for home appliances, if the network is expanded based on multiple dimensions such as brand, category, function, and parameter, the resulting network hierarchy may exceed the preset threshold of five levels. In this case, it is necessary to merge some similar functional features, such as combining cooling power and power consumption into energy efficiency indicators, to optimize the network structure. Weight assignment should take into account the importance of features. For example, in mobile phone products, price range and performance configuration should be weighted higher than appearance and color. By establishing a weight assignment table, core features such as processor performance can be assigned a weight of 0.8, while appearance and color can be weighted 0.3. A dynamic programming algorithm adaptively adjusts the connectivity of feature nodes based on the weight assignment table. In the optimized association graph, the connections between feature nodes are more rational. This weighted connectivity more accurately reflects the degree of correlation between features, which facilitates subsequent product recommendations and marketing strategy development. This multi-level feature analysis method can better understand the correlations between product features, providing data support for precision marketing and personalized recommendations. The optimized association graph can quickly locate key feature nodes, improving the efficiency and accuracy of feature analysis.

[0035] Step S108 : Based on the optimized feature weight distribution scheme, the multi-path feature combination with the highest weight is selected from the feature association graph to generate interest-based recommendation content.

[0036] Based on the optimized network depth and connection density, feature weights are reallocated to generate a weight distribution table. A dynamic programming algorithm, combined with the weight distribution table, is used to update feature weights in the association graph. Using the updated feature weights, the multi-path feature combinations with the highest weights are extracted. Based on these extracted multi-path feature combinations, recommendations are generated for specific interest groups.

[0037] Specifically, feature weight assignment is a key step in optimizing the association graph. Weights are reallocated to reflect the importance of different features. The weight assignment table also considers market factors such as price range and brand positioning, increasing the brand weight of high-end brands to above 0.7. Dynamic programming algorithms iteratively update feature weights. In the apparel sector, the weights of features such as fabric texture, style, design, and brand level adjust seasonally. In summer, the weight of fabric breathability increases to 0.8, while warmth decreases to 0.3. The algorithm automatically adjusts weights based on sales data. For example, if a dress becomes popular due to its unique cut, the weight of the design feature will be increased accordingly. Extracting multi-path feature combinations requires a comprehensive consideration of the strength of the associations between features. For example, for beauty products, active ingredients, skin type, and price positioning constitute a highly weighted feature combination. Hyaluronic acid and vitamins, due to their high co-occurrence among active ingredients, have a weight of 0.85. The algorithm identifies the path with the highest weight for the active ingredient plus skin type combination and makes subsequent recommendations based on this weight. Personalized recommendations based on feature combinations must adapt to changing user interests. In household items, decoration style, space size, and color matching form a typical feature combination. If the system finds that the user has recently browsed Nordic-style furniture, it will extract a high-weight feature combination that includes simple design, solid wood material, and gray and white color matching. When a user clicks on a certain feature combination multiple times, the recommendation weight of this combination will be further improved. Recommendations in the food and beverage field pay more attention to the balance between taste and nutrition. After identifying the user's preference for a low-sugar, high-protein feature combination, the system will give priority to recommending healthy snacks that meet this feature. If a user frequently purchases a specific flavor of a certain brand, the weight of this feature combination will increase accordingly. Through dynamically updated weight distribution, the recommendation system can respond to shifts in user interests in a timely manner and provide more accurate personalized content.

[0038] Step S109, matching the generated personalized recommendation content with the user's real-time feedback behavior, adjusting the recommendation strategy, and continuously improving the recommendation conversion rate. The user's real-time feedback behavior specifically refers to the user's retention time and sharing and forwarding behavior.

[0039] Obtain user retention time and sharing and forwarding behavior data, and calculate the degree of match between the number of behaviors and the feature quantity. Dynamically adjust the recommendation strategy value based on the match, and update the weight of personalized recommended content. Use a dynamic programming algorithm combined with the weight value to recalculate the relationship between the number of recommendations and the conversion rate. If the conversion rate falls below the preset threshold, optimize the combination of feature quantity and strategy value. Based on the optimized feature combination, generate new personalized recommendations. Use real-time behavioral data to determine the degree of match between recommended content and user behavior. If the match improves, determine that recommendation strategy as the final optimization solution.

[0040] Specifically, user retention length reflects content attractiveness and user stickiness, and sharing and forwarding reflect social communication value. By tracking the user's stay time and interaction behavior on the platform, the feature matching degree can be calculated. For example, the fashion shopping platform finds that the average browsing time of users for a product is more than ten minutes, and the matching degree of the style features of the product and the user portrait is usually above 0.8. The ingredients and efficacy features of high-frequency sharing and forwarding beauty products are often highly consistent with the user's skin care needs. Dynamically adjusting the recommendation strategy needs to balance the recommendation accuracy and content diversity. If it is found that the average stay time of users for a certain category decreases, the system will accordingly reduce the recommendation weight of that category. If users frequently share sports equipment related content to their friends circle, 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 user's interest changes. The relationship between the recommended amount and the conversion rate reflects the recommendation effect. Taking maternal and infant products as an example, when the system recommends twenty items of goods per day, the conversion rate is 5%. After optimization by dynamic programming algorithm, the recommended amount of high matching degree goods is increased to thirty, and the recommended amount of low matching degree goods is reduced to ten, and the conversion rate is increased to 8%. This shows that precise targeted recommendation can improve the user's purchase willingness. The combination of feature quantity and strategy value determines the recommendation effect. In the field of home appliances, the weight proportion of features such as price range, function configuration, and brand level will affect user decision-making. When it is found that the conversion rate of high-priced brand refrigerators is lower than expected, the system will adjust the feature combination, increase the weight of energy-saving rating, user experience and other practical features, and improve the recommendation effect. Real-time behavior data reflects the user's acceptance of recommended content. In book recommendation, the reading time, note quantity and other behavior indicators of users after obtaining recommendations can directly reflect the matching degree. If it is found that the reading completion rate of users for popular science books is improved, and they often add bookmarks and make notes, it means that the recommendation strategy of this type of books has high matching degree. The system will solidify this strategy as an optimization scheme and continue to use and fine-tune it in subsequent recommendations. The recommendation of fresh food needs to consider the seasonal and local features. When the in-season fruits get higher stay time and sharing amount, the system will increase the recommendation weight of similar categories. At the same time, according to the user's purchase frequency and evaluation, the combination of features such as origin and price range is dynamically adjusted to continuously optimize the recommendation effect. By continuously monitoring the user's feedback on different feature combinations, the recommendation system can establish a more accurate personalized service model.

[0041] In step S1010, the feature weight distribution in multiple time scales is fused to determine the time period mode in which the user is currently located, and the weight proportion is adjusted. If the user is in the rush hour period of the working day, the search hot words and real-time click feedback in the last hour are focused on, and the weight proportion of these features is increased. If the user is in the leisure period at night on the weekend, the topic preference and video type distribution in the past week are focused on, and the weight proportion of these features is increased.

[0042] Obtain user behavior data at different time scales, including search hot words, click feedback, topic preferences, and video type distribution. Use pattern recognition algorithms to determine the time period patterns of users during peak commuting hours or late-night leisure time. Based on the time period judgment results, dynamically adjust the weight of feature distribution to increase the weight of behavioral data for the corresponding time period. Use clustering algorithms to classify user behavior data and generate feature combinations for different time periods. Based on these feature combinations, recalculate user interests and preferences for different time periods. Use collaborative filtering algorithms to combine user interests and behavior data to generate personalized recommendations. Dynamically update the weight of recommended content based on real-time user behavior data to optimize recommendation strategies.

[0043] Specifically, time-scale user behavior data collection is crucial for recommendation systems. During the commute period, urban users tend to search for breakfast restaurants, weather forecasts, and other practical information, and watch short videos with a duration of three to five minutes. During the late-night leisure period, users spend more time browsing entertainment content such as food and travel, with an average viewing time of more than fifteen minutes. Pattern recognition algorithms can analyze user activity patterns to determine time periods. For example, on a certain video platform, during the early morning hours of weekdays, the search keywords of the commuter group are mainly concentrated in the fields of news information and traffic conditions. After 10 pm, the click-through rate of leisure and entertainment content significantly increases, and users' acceptance of long videos also increases. 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 weight of light content such as food and entertainment can be reduced to 0.2. This adjustment ensures that the pushed content matches the user's current time period needs. Behavior data clustering analysis can reveal the time period characteristics of users. A news information platform found that office workers tend to quickly browse headlines in the morning, prefer to watch light life topics during lunch breaks, and prefer in-depth reading and knowledge content on weekends. Based on these characteristics, the platform can push more matching content during different time periods. The calculation of user interest preferences needs to consider the time dimension. For example, during the morning commute 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 at night, the recommendation weight of food content will be increased to 0.5 accordingly. Collaborative filtering algorithms play a key role in personalized recommendations. Education platform data shows that the concentration of high school students watching teaching videos during the evening study period is high, and the conversion rate of providing similar course recommendations can reach 30%. On weekends, the same group has a higher acceptance of interest development content, and personalized recommendations need to be adjusted in time. Real-time behavior data updates ensure that recommended content remains fresh. Music platforms have found that users have different music style preferences at different times. In the morning, they tend to prefer light music, accounting for 40%, while at night they prefer songs with strong rhythm, accounting for 60%. By adjusting the recommendation weight of music type in real time, the user's stay time can be significantly improved.

[0044] As Figure 4As shown, an 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 acquisition module, which is used to obtain multi-dimensional behavioral data of users in scenarios such as information browsing and short video watching, and extract original features such as search terms, browsing time, likes and comments; a feature extraction module, which is used to construct behavioral feature views of different time granularities such as seconds, minutes, and hours for the extracted original features, and form a multi-time scale feature representation; a time granularity view construction module, which is used to calculate the time weight distribution of feature nodes in the feature view of each time granularity, and use an exponential decay function to perform time decay processing on historical features; a time weight calculation module, which is used to analyze feature views at different time scales through a graph mining algorithm, mine association rules between feature nodes, and construct a cross-scale feature association graph; a graph mining analysis module, which is used to judge the time correlation of different feature combinations in the feature association graph based on the spectral clustering results of the feature nodes and the gravitational direction of the feature edges; a feature association graph construction module, which is used to assign a high weight coefficient of a power-law distribution to the time weight of the feature combination if the time correlation of the feature combination exceeds the empirical threshold; The time correlation judgment module is used to dynamically adjust the network depth and connection density of the association graph based on the vertical dimension and semantic level of the feature, and optimize the feature weight distribution in the graph; the weight optimization module is 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 to generate interest-based recommended content; the recommended content generation module is used to 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 recommendations; the real-time feedback adjustment module is used to integrate the feature weight distribution at multiple time scales, determine the user's current time cycle pattern, and adjust the weight ratio. If the user is in the weekday commuting peak period, the focus is on their search hot words and real-time click feedback in the last hour, and the weight ratio of these features is increased. If the user is in the late-night leisure time on the weekend, the focus is on their topic preferences and video type distribution within the week, and the weight ratio of these features is increased. Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled 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 advertising push method based on multi-dimensional historical data analysis, characterized in that: The method comprises: Acquire multi-dimensional user behavior data in different scenarios and extract raw feature data. Based on the extracted raw features, construct behavioral feature views at different time granularities, such as seconds, minutes, and hours, to form a multi-time-scale feature representation. In the feature view at each time granularity, calculate the time weight distribution of feature nodes, and use an exponential decay function to perform time decay on historical features. Analyze the feature views at different time scales using graph mining algorithms, discover association rules between feature nodes, and construct a cross-scale feature association graph. In the feature association graph, determine the temporal correlation of different feature combinations based on the spectral clustering results of feature nodes and the gravitational direction of feature edges. If the temporal correlation of a feature combination exceeds an empirical threshold, assign a high power-law-distributed temporal weight to the feature combination. Dynamically adjust the network depth and connection density of the association graph based on the vertical dimension and semantic level of the features to 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 interest-based recommendations. Match the generated personalized recommendations with the user's real-time feedback behavior, adjust the recommendation strategy, and continuously improve the conversion rate of recommendations. In the feature association graph, the temporal correlation of different feature combinations is determined based on the spectral clustering results of the feature nodes and the gravitational direction of the feature edges, including: 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 value; according to the gravitational direction of the feature edge, the combination point relationship between different feature combinations is calculated to obtain the gravitational strength of the combination point; the time degree distribution between different feature groups is determined by combining the grouping value and the combination point relationship to generate the time correlation degree; if the time correlation degree is higher than the preset threshold, the feature group is marked as a high-correlation combination, and the feature nodes in the high-correlation combination are extracted; according to the feature nodes of the high-correlation combination, the time granularity priority in the feature view is adjusted to generate an optimized feature view; using the optimized feature view, the gravitational direction of the feature edge is recalculated and the gravitational strength in the feature association graph is updated; through the updated feature association graph, the spectral clustering algorithm parameters are optimized to improve the accuracy of feature node clustering; The method of dynamically adjusting the network depth and connection density of the association graph based on the vertical dimension and semantic level of the features and optimizing the feature weight distribution in the graph includes: Based on the vertical dimension and semantic level of the features, an initial association graph is constructed; a hierarchical clustering algorithm is used to stratify the graph nodes to obtain the node distribution at different levels; based on the node level distribution, the network depth and connection density of the graph are calculated; if the network depth exceeds the preset threshold, the hierarchical structure of the graph is adjusted to optimize the connection density; based on the optimized network depth and connection density, the feature weights are reallocated to generate a weight distribution table; a dynamic programming algorithm is used in combination with the weight distribution table to update the feature weights in the association graph; based on the updated feature weights, the node connection relationship of the graph is adjusted to generate an optimized association graph.

2. The method according to claim 1, characterized in that The acquisition of multi-dimensional behavior data of users in different scenarios and the extraction of original feature data include: Obtaining behavioral data of at least one user in information browsing and short video viewing scenarios, wherein the behavioral data includes 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; The 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 a composite feature. If the contribution of a feature in the behavioral data to the target is lower than a preset threshold, the feature is eliminated.

3. The method according to claim 1, characterized in that The extracted original features are used to construct behavioral feature views at different time granularities, such as seconds, minutes, and hours, to form multi-time scale feature representations, including: Obtain user behavioral data in information browsing and short video viewing scenarios, and construct behavioral feature views with multiple time granularities at the second, minute, and hour levels 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.

4. The method according to claim 3, characterized in that 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, including: Construct behavioral feature views at multiple time granularities, such as seconds, minutes, and hours, and use an exponential decay function to calculate the time weight distribution of feature nodes at each time granularity. According to the preset attenuation coefficient, the historical features of the second-level, minute-level, and hour-level feature nodes are subjected to time decay processing to obtain the attenuated feature values; For the attenuated second-level, minute-level, and hour-level feature values, a weighted average algorithm is used to calculate the comprehensive score of each feature.

5. The method according to claim 4, characterized in that The graph mining algorithm is used to analyze feature views at different time scales, mine association rules between feature nodes, and construct a cross-scale feature association graph, including: Graph mining algorithms are used to analyze feature views at the second, minute, and hour levels to obtain association rules between feature nodes. Based on the mined association rules, a cross-scale feature association graph is constructed to describe the association relationships between feature nodes. If there are strongly associated nodes in the feature association graph, they will be marked as key feature nodes; Adjust the time granularity priority in the feature view according to the weight distribution of key feature nodes; Through the adjusted feature views, the graph mining algorithm parameters are optimized and the accuracy of association rule mining is improved.

6. The method according to claim 1, characterized in that If the time correlation of the feature combination exceeds the empirical threshold, the time weight of the feature combination is assigned a high weight coefficient of the power law distribution, including: The spectral clustering algorithm is used to group the node set and obtain the clustering results of the characteristic nodes; According to the gravitational direction of the feature edge, the relationship between the combination points of different feature combinations is calculated to obtain the gravitational strength; Combining the clustering results and the gravity strength, the temporal correlation of the feature groups is determined and the temporal degree distribution is generated; If the time correlation exceeds the preset value, a power law distribution algorithm is used to assign a high weight coefficient to the time weight; According to the high weight coefficient, the time granularity priority in the feature view is adjusted to generate an optimized feature view.

7. The method according to claim 1, characterized in that The method of 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-based recommendation content includes: Redistribute feature weights based on the optimized network depth and connection density to generate a weight distribution table; Use dynamic programming algorithm combined with weight distribution table to update feature weights in association graph; Extract the multi-path feature combination with the highest weight through the updated feature weights; Generate interest-based recommendation content based on the extracted multi-path feature combination.

8. The method according to claim 1, characterized in that The generated personalized recommendation content is matched with the user's real-time feedback behavior, the recommendation strategy is adjusted, and the conversion rate of the recommendation is continuously improved. The user's real-time feedback behavior is specifically the user's retention time and sharing and forwarding behavior, including: Obtain user retention time and sharing and forwarding behavior data, and calculate the matching degree between behavior data and feature quantity; Dynamically adjust the recommendation strategy value based on the matching degree and update the weight value of personalized recommendation content; Use dynamic programming algorithm combined with weight value to recalculate the relationship between recommendation volume and conversion rate; If the conversion rate is lower than the preset threshold, the combination of feature quantity and strategy value is optimized; Generate new personalized recommendation content based on the optimized feature combination; Use real-time behavioral data to determine the matching degree between recommended content and user behavior; If the matching degree is improved, the recommended strategy is determined as the final optimization solution.

Images