Internet advertisement intelligent promotion method and system

Abstract

The invention belongs to the technical field of business data processing, and particularly discloses an internet advertisement intelligent promotion method and system.The method comprises the steps that data such as browsing duration and rolling speed of a user in the advertisement interaction process are collected, a behavior sequence is obtained to reflect performance differences in different scenes, and multi-scale behavior records are obtained; processing continuous interaction in a time dimension by adopting a sequence model according to the acquired behavior sequence, extracting characteristics such as short-term sight line staying and long-term habits, and determining multi-layer behavior representation; aiming at the determined multi-layer behavior representation, highlighting correlation between key interaction points such as a page stay peak value and a rolling frequency by applying an attention mechanism, judging weight distribution among characteristics, and obtaining a weighted behavior vector; if the short-term characteristic dominance in the obtained weighted behavior vector exceeds a preset threshold value, vector balance is adjusted by fusing long-term habit data; the objective of the invention is to solve the problem that user experience is affected by poor pertinence of Internet advertisement recommendation in the prior art.

Description

Technical Field

[0001] This invention relates to the field of commercial data processing technology, specifically to an intelligent promotion method and system for internet advertising. Background Technology

[0002] Intelligent internet advertising promotion, as a core direction in digital marketing, holds undeniable value in improving advertising effectiveness and user experience. With the increasing complexity and diversity of user behavior, accurately capturing user intent and achieving personalized advertising recommendations has become a crucial issue for the industry. Research in this area not only optimizes advertising effectiveness but also directly impacts business revenue and users' online experience. However, many current methods, when processing user behavior data, often lack in-depth mining and detailed analysis of behavioral characteristics, making it difficult to fully understand users' true needs in different scenarios. Especially when facing the complex process of user-advertisement interaction, existing technologies often overlook the multi-layered information hidden behind behavioral data, leading to biased judgments of user intent and consequently affecting the accuracy of advertising recommendations. This limitation makes the system unable to adapt to dynamic changes in user behavior and unable to cope with personalized needs in different scenarios.

[0003] Against this backdrop, the research field faces significant technical challenges, the most critical of which lies in how to refine and decompose user behavior data at multiple levels. User behavior when browsing advertisements encompasses a wide range, from subtle physiological responses to overall interaction habits. The differences in these behavioral characteristics across different temporal and spatial scales directly lead to the system's difficulty in accurately capturing subtle changes in user intent. Furthermore, this variability presents another challenge: how to integrate these behavioral characteristics at different scales into a unified analytical framework. For example, when users browse an advertisement page, a brief glance might reflect interest in a particular section of content, while frequent page scrolling might suggest rapid filtering of the overall content. Without combining these subtle behaviors with overall interaction habits for analysis, the system struggles to determine the user's true attitude towards the advertisement.

[0004] Therefore, how to accurately analyze changes in user intent at different scales based on multi-level behavioral characteristics, and apply these analysis results to the precise matching of advertising recommendations, has become a key problem that this study urgently needs to solve. Summary of the Invention

[0005] This invention provides an intelligent promotion method and system for internet advertising, aiming to solve the problem of poor targeting of existing internet advertising recommendations, which affects user experience.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for intelligent internet advertising promotion includes: collecting data such as browsing time and scrolling speed during user interaction with advertisements to obtain behavioral sequences that reflect performance differences in different scenarios, resulting in multi-scale behavioral records; processing continuous interactions over time using a sequence model based on the obtained behavioral sequences to extract features such as short-term gaze dwell and long-term habits, and determining multi-layered behavioral representations; applying an attention mechanism to highlight key interaction points, such as the correlation between peak page dwell time and scrolling frequency, and determining the weight distribution among features to obtain a weighted behavioral vector; if the short-term features dominate the obtained weighted behavioral vector beyond a preset threshold, adjusting the vector balance by fusing long-term habit data to obtain an integrated behavioral description; extracting intent change patterns, such as interest shift trajectories, from the obtained integrated behavioral description to determine the user's potential preferences for advertisement content and determining an intent vector; matching the most similar items in a pre-established advertisement library based on the determined intent vector to obtain a preliminary recommendation list; dynamically sorting the obtained preliminary recommendation list using a sequence model, adjusting priorities based on real-time user feedback, determining the final output order, and obtaining a personalized advertisement sequence.

[0007] In one aspect of this disclosure, by collecting data such as user browsing time and scrolling speed during advertising interaction to obtain behavioral sequences to reflect performance differences in different scenarios, multi-scale behavioral records are obtained, including: By capturing users' browsing time and scrolling speed during ad interactions, initial behavioral sequence data is constructed; Using pre-established data collection tools, relevant information is extracted from interaction records to obtain a preliminary user behavior dataset; Based on the preliminary user behavior dataset, we analyzed the distribution of browsing time and scrolling speed in different scenarios. For duration distribution and speed changes, classification methods are applied to determine the behavioral pattern categories under different scenarios; if the duration distribution of the classified behavioral patterns deviates from the preset threshold in a certain scenario, the data in that scenario is then deeply mined. By further breaking down the performance differences, we can obtain more granular multi-scale data records.

[0008] In one aspect of this disclosure, the step of processing continuous interactions over time using a sequence model based on the acquired behavioral sequence, extracting features such as short-term gaze dwell and long-term habits, and determining multi-layered behavioral representations includes: By obtaining behavioral sequences from user interactions, we can initially break down continuous behaviors over time to obtain short-term and long-term behavioral fragments. Based on short-term and long-term behavioral segments, a sequence model is used to analyze the duration of gaze fixation and identify significant changes in short-term characteristics. For significant changes in short-term characteristics, repetitive patterns in long-term behavioral habits are obtained to determine the stability of long-term characteristics. If the stability of long-term characteristics is lower than a preset threshold, the behavioral segments are regrouped. Based on the regrouped behavioral fragments, key behavioral nodes are extracted from the multi-layered representation to obtain the core behavioral patterns in user interaction. For core behavioral patterns, analyze the distribution of behavior over time to identify potential correlations in continuous behaviors; By identifying potential association points, a multi-layered behavioral representation of user interaction is constructed, resulting in the final set of behavioral features. Based on the set of behavioral features, a behavioral change trajectory over time is generated to determine the behavioral preferences in user interactions.

[0009] In one aspect of this disclosure, the method of applying an attention mechanism to highlight key interaction points, such as the correlation between peak page dwell time and scroll frequency, for a given multi-layer behavior representation, determining the weight distribution among features, and obtaining a weighted behavior vector includes: For key nodes in the behavioral representation, an attention mechanism is used to process the time segments of user operations in layers, identify the correlation pattern between page dwell time and scrolling frequency, and obtain a preliminary distribution of interaction priorities. Based on the initial distribution of interaction focus, analyze the peak positions of page dwell time, and combine the changes in scrolling frequency to determine the correlation strength between peak dwell time and frequency. To determine the correlation strength between peak dwell times and frequency, feature weights are calculated, and the weight distribution is sorted to identify high-priority interaction areas. By identifying high-priority interactive key areas, the core time segments in user operations are extracted. If the correlation analysis results of the core time segments are lower than the preset threshold, the segments are re-divided to obtain an adjusted set of time segments. Based on the adjusted set of time segments, a weighted vector is constructed, and the behavior representation is updated for the high-value regions in the weight distribution to determine the final weighted behavior pattern. By analyzing the final weighted behavior pattern, the performance of key nodes in different time segments is analyzed to determine the continuity characteristics of user operations and obtain complete interaction analysis results.

[0010] In one aspect of this disclosure, if the short-term features dominate more than a preset threshold in the obtained weighted behavior vector, the vector balance is adjusted by fusing long-term habit data to obtain an integrated behavior description, including: For short-term behavioral characteristics in the weighted behavioral vector, the proportion analysis is used to determine whether they exceed the preset threshold. If the proportion of short-term behaviors exceeds a preset threshold, long-term habit data will be obtained from a pre-established behavior database to determine the initial direction of adjustment. Based on the initial adjustment direction, a data introduction approach is adopted to combine long-term habit data with short-term behavioral characteristics to obtain a mixed behavioral dataset, which will then be used as an intermediate result for subsequent processing. For the mixed behavior dataset in the intermediate results, the proportion of short-term behavior and long-term habit in the behavior vector is redistributed through a balancing adjustment method to determine the adjusted vector structure; Based on the adjusted vector structure, analyze the changes in the proportion of behaviors, obtain the updated behavior vectors, and get an intermediate description that meets the balance requirements; For the update vector in the intermediate description, the characteristics of short-term behavior and long-term habits are integrated through a comprehensive description construction method to determine the final behavior representation. Based on the final behavioral representation, the completeness of the results after vector update is analyzed and adjusted to obtain a comprehensive behavioral description.

[0011] In one aspect of this disclosure, the step of obtaining intent change patterns, such as interest shift trajectories, from the obtained integrated behavioral description, determining the user's potential preference for advertising content, and identifying the intent vector includes: Data related to changes in intent are obtained from behavioral descriptions. The change trajectory is initially extracted, and time series processing is used to sort out the changes in user patterns and obtain preliminary trajectory records. Based on the initial trajectory records, the specific direction of interest shifts is analyzed, and the data is split according to the shift pattern. By segmenting the data, the distribution of interest points in different time periods is identified, and the segmentation results of the shift pattern are determined. Based on the segmentation results of the transfer mode and combined with relevant data of the advertising content, the matching of content preferences is analyzed. By using a comparative processing method, the differences in user responses to different advertising content are identified, and a preliminary judgment on preference matching is obtained. Based on the initial judgment of preference matching, relevant information on potential tendencies is extracted. If the intensity of the potential tendency response exceeds the preset threshold, the priority of the tendency judgment is adjusted through weighted processing to determine the intermediate data of the tendency judgment. Based on the intermediate data of tendency judgment, combined with the comprehensive information of change trajectory and transfer pattern, the basic structure of intention vector is constructed by using logical mapping, and a preliminary representation of intention vector is obtained. Based on the initial representation of the intent vector, the data is integrated according to the analysis results. Through multi-dimensional verification, the structure of the intent vector is refined to determine the final form of the intent vector.

[0012] In one aspect of this disclosure, the step of matching the most similar entries in a pre-established ad library based on a determined intent vector to obtain a preliminary recommendation list includes: For the initial recommendation set, the item data with the highest correlation with the intent vector is obtained. A hierarchical comparison method is used to sort out the correlation strength between the item and the vector, and a priority-ranked list of items is obtained. Based on the priority-sorted list of items, obtain the top-ranked ad resources in the list, use content parsing tools to extract the theme information of the ad resources, and determine the fit data between the theme and the intent vector; If the relevance data between the theme and the intent vector exceeds a preset threshold, the corresponding ad resource entry will be retained. If the matching data is lower than the preset threshold, the entry is removed, and a filtered subset of resources is obtained. Based on the selected resource subset, the content-related information of each item in the subset is obtained. Through multi-angle comparison, the complementarity between items is analyzed to determine the combination of items with high complementarity. For combinations of items with high complementarity, the distribution of advertising resources within the combination is obtained. If the resource distribution within the combination is uneven, the combination structure is adjusted by re-sorting to obtain an optimized recommended combination. Based on the optimized recommended combinations, the final item data within the combinations is obtained. Logical verification is used to confirm the matching accuracy between the items and the intent vector, and the final recommendation result is determined.

[0013] In one aspect of this disclosure, the method of dynamically sorting the obtained preliminary recommendation list using a sequence model, adjusting priorities based on real-time user feedback, determining the final output order, and obtaining a personalized advertising sequence includes: A sequence model is used to perform initial sorting on the preliminary recommendation list, and the relative position data between items is obtained to determine the preliminary sorting result. Based on the preliminary ranking results, real-time user feedback data is obtained. By analyzing the behavioral preference information in the feedback, if the feedback data indicates that users show high attention to certain items, the priority of the corresponding items is increased, resulting in an adjusted ranking set. For the adjusted sorted set, obtain the priority data of each item in the set. By comparing the priorities, if the priority of an item is lower than the preset threshold, move its position to the right to determine the optimized order of items. Based on the optimized item order, obtain the data of the top-ranked ad items in the order, use content parsing tools to extract the core information of the items, and determine the set of ad content related to user preferences; For the set of advertising content, obtain the correlation data between each item in the set. Through multi-angle comparison, if some items are found to have overlapping content, the duplicate items are removed to obtain a simplified advertising sequence. Based on the streamlined ad sequence, the distribution of the remaining items in the sequence is obtained. Through logical verification, if an item deviates from the theme of the overall sequence, its position is adjusted to determine the final personalized recommendation order. For the final personalized recommendation order, the display information of each item in the order is obtained, and a complete advertising presentation plan is generated through data integration tools to determine the final output advertising sequence.

[0014] In another aspect, this disclosure also relates to an intelligent internet advertising promotion system, comprising: The data acquisition module is used to collect data such as browsing time and scrolling speed during the user's advertising interaction process, obtain behavioral sequences to reflect the performance differences in different scenarios, and obtain multi-scale behavioral records. The sequence processing module is used to process continuous interactions in the time dimension using a sequence model based on the acquired behavioral sequences, extract features such as short-term gaze lingering and long-term habits, and determine multi-layered behavioral representations. The feature extraction module is used to apply an attention mechanism to highlight key interaction points, such as the correlation between peak page dwell time and scroll frequency, for a given multi-layer behavior representation, determine the weight distribution between features, and obtain a weighted behavior vector. The attention weighting module is used to adjust the vector balance by fusing long-term habit data if the short-term features dominate the obtained weighted behavior vector more than a preset threshold, so as to obtain an integrated behavior description. The vector integration module is used to extract intent change patterns, such as interest shift trajectories, from the obtained integrated behavior descriptions, determine users' potential preferences for advertising content, and identify intent vectors. The intent analysis module is used to match the most similar items in a pre-built ad library based on the determined intent vector to obtain a preliminary recommendation list; The recommendation ranking module uses a sequence model to dynamically rank the initial recommendation list, adjusts the priority based on real-time user feedback, determines the final output order, and obtains a personalized ad sequence.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a multi-layered behavioral representation by deeply analyzing data such as browsing time and scrolling speed, and integrates an attention mechanism to highlight key interaction points, forming a weighted behavioral vector. When short-term features dominate, long-term habit data is used to adjust the balance, generating an integrated behavioral description, accurately capturing the user's interest shift trajectory, and determining the intent vector. This invention further optimizes the recommendation list by matching the ad library and dynamic ranking, combined with real-time feedback, ultimately outputting a personalized ad sequence. This invention, through the combination of sequence models and attention mechanisms, achieves intelligent processing from behavioral data to intent judgment to precise recommendation, effectively improving the targeting and user experience of ad recommendations, and demonstrating the technical advantages of deeply mining user preferences in dynamic interaction scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an intelligent internet advertising promotion method according to the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0019] Please see Figure 1 As shown in the figure, this embodiment discloses an intelligent promotion method and system for internet advertising, which may specifically include: Step 101: By collecting data such as browsing time and scrolling speed of users during the advertising interaction process, behavioral sequences are obtained to reflect the performance differences in different scenarios, and multi-scale behavioral records are obtained.

[0020] Initial behavioral sequence data is constructed by capturing user browsing duration and scrolling speed during ad interactions. Using pre-established data collection tools, relevant information is extracted from interaction records to obtain a preliminary user behavior dataset. Based on this dataset, the distribution of browsing duration and scrolling speed across different scenarios is analyzed. Classification methods are applied to determine behavioral pattern categories under different scenario variations, based on duration distribution and speed changes. If the duration distribution of the classified behavioral patterns deviates from a preset threshold in a particular scenario, in-depth data mining is performed on the data for that scenario. Further breakdown of performance differences yields finer-grained multi-scale data records. Key change points in user behavior during ad interactions are extracted from these multi-scale data records. Support Vector Machine (SVM) algorithms are used to perform pattern matching on these key change points to determine potential correlations within the behavioral sequences. If the correlation in the behavioral sequences is below a preset threshold, the interaction records undergo secondary processing. A more precise user behavior feature set is obtained by recombining the duration distribution and speed changes. Continuous analysis of the user behavior feature set constructs a performance difference mapping under different scenarios. Based on the mapping results, the final multi-scale data integration results are obtained to determine the direction for behavioral optimization in ad interactions.

[0021] Specifically, in the process of collecting and analyzing user advertising interaction behavior data, the process begins by using front-end tracking technology to record the user's browsing time and scrolling speed on the advertising page in real time. For example, JavaScript is used to capture the user's entry and exit times, and the browsing time is calculated. Assuming a user stayed on ad A for 45.6 seconds, the scrolling speed is recorded by listening to page scrolling events, resulting in an average scrolling pixel value of 30.5 pixels per second. This data is stored as a timestamp sequence, forming a preliminary behavioral data stream. Next, this data is divided into behavioral sequences according to time windows. For example, using 5-second windows, the 45.6-second browsing time is divided into 9 windows. Within each window, the average scrolling speed and browsing depth are calculated to obtain a serialized feature vector. A K-means clustering algorithm (with K=3) is used to classify the behavioral features of different windows, analyzing the user's behavioral patterns in different time periods. It was found that the first three windows had higher scrolling speeds, with an average of 35.2 pixels per second, indicating stronger initial user interest. Furthermore, multi-scale behavioral records were constructed based on these sequences. Wavelet transform algorithms were used to decompose the behavioral sequences at multiple scales, extracting low-frequency and high-frequency signals. It was assumed that low-frequency signals reflected the overall browsing trend of users, while high-frequency signals captured instantaneous behavioral fluctuations. Analysis revealed that the low-frequency trend of users on ad A indicated sustained attention, while the high-frequency fluctuation peaked at the 20-second mark, possibly related to key trigger points in the ad content. Finally, by comparing multi-scale records under different scenarios, such as user behavior data from ad A and ad B, it was found that the average browsing time for ad A was 45.6 seconds, higher than ad B's 32.8 seconds, and the scrolling speed fluctuations were more stable. Combined with the business scenario, it was inferred that the content design of ad A was more attractive to users, thus optimizing the ad placement strategy. All data processing was automated through the backend server, forming a closed-loop logic from collection to analysis.

[0022] Step 102: Based on the acquired behavioral sequence, use a sequence model to process continuous interactions in the time dimension, extract features such as short-term gaze lingering and long-term habits, and determine multi-layered behavioral representations.

[0023] By analyzing behavioral sequences obtained from user interactions, continuous behaviors over time are initially broken down to obtain short-term and long-term behavioral segments. Based on these segments, a sequence model is used to analyze the duration of gaze lingering, identifying significant changes in short-term characteristics. For these significant changes, repetitive patterns in long-term behavioral habits are extracted to assess their stability. If the stability of long-term characteristics falls below a preset threshold, the behavioral segments are regrouped. Based on the regrouped segments, key behavioral nodes are extracted from the multi-layered representation to obtain core behavioral patterns in user interactions. For these core patterns, the distribution of behavior over time is analyzed to identify potential correlations in continuous behaviors. Using these potential correlations, a multi-layered behavioral representation of user interactions is constructed, resulting in a final set of behavioral features. Based on this set, a behavioral change trajectory over time is generated to determine user behavioral preferences.

[0024] Specifically, in processing user advertising interaction data to construct multi-layered behavioral representations, a series of automated information technology methods can be used to analyze continuous interactions over time. First, for the acquired behavioral sequences, the system uses time series segmentation technology to divide the user's interaction data on the advertising page into fixed time intervals, for example, 60 seconds of interaction data divided into 6 time segments of 10 seconds each. The system extracts the gaze duration for each segment. Assuming a user's gaze duration is 6.2 seconds in the first segment and only 3.1 seconds in the last segment, this initially reflects a decreasing trend in user attention over time. Next, a recurrent neural network algorithm is used to model the continuous interaction data of these time segments. With a hidden layer dimension of 64, the model is trained to capture short-term gaze duration characteristics. Analysis shows that the average short-term gaze duration reaches 5.8 seconds within the initial 30 seconds, indicating that users pay higher attention to the first half of the advertising content. Subsequently, the system further extracts long-term habitual features using a Long Short-Term Memory (LSTM) network. With a time step of 5, it analyzes user interaction patterns across multiple ads, revealing an average interaction duration of 48.7 seconds for similar ads over the past week, reflecting long-term preferences. Finally, by weighting and fusing short-term and long-term features with weights of 0.4 and 0.6 respectively, a multi-layered behavioral representation is constructed. Analysis shows a comprehensive behavioral tendency value of 0.72 for the current ad, which can be used to adjust the presentation order of ad content based on business needs. All processing is completed by a backend automated system, forming a complete logical chain from data extraction to feature fusion.

[0025] Step 103: For the determined multi-layer behavior representation, apply the attention mechanism to highlight the correlation between key interaction points such as the peak page dwell time and scrolling frequency, determine the weight distribution among features, and obtain the weighted behavior vector.

[0026] For key nodes in the behavioral representation, an attention mechanism is used to layer user operation time segments, identifying the correlation patterns between page dwell time and scrolling frequency to obtain a preliminary distribution of interaction priorities. Based on this preliminary distribution, the peak positions of page dwell time are analyzed, and the correlation strength between peak dwell time and frequency is determined by combining this with changes in scrolling frequency. Feature weights are calculated for the correlation strength between peak dwell time and frequency, and the weight distribution is sorted to obtain high-priority interaction priority areas. Core time segments in user operations are extracted from these high-priority interaction priority areas. If the correlation analysis results of core time segments are below a preset threshold, the segments are re-divided to obtain an adjusted set of time segments. Based on the adjusted set of time segments, a weighted vector is constructed, and behavioral representations are updated for high-value areas in the weight distribution to determine the final weighted behavioral pattern. Using the final weighted behavioral pattern, the performance of key nodes in different time segments is analyzed to determine the persistence characteristics of user operations, resulting in a complete interaction analysis.

[0027] Specifically, in processing user ad interaction data to construct a weighted behavior vector, the system uses automated information technology to perform in-depth analysis of multi-layered behavior representations, highlighting key interaction points and optimizing feature weight distribution. First, the system extracts detailed user interaction data on the ad page from the constructed multi-layered behavior representation, such as peak page dwell time and scrolling frequency. Assuming a user's peak dwell time on an ad page reaches 12.5 seconds and the scrolling frequency is 8.3 times per minute, the system standardizes these indicators through a data preprocessing module, forming a unified input vector. Next, the system applies an attention mechanism algorithm, setting the attention layer dimension to 32. By calculating the attention score for each interaction point, the analysis reveals a correlation of 0.85 between peak page dwell time and scrolling frequency, indicating a strong correlation in user interest judgment. Subsequently, the system adjusts the feature weights based on the attention scores. Assuming the initial weight for peak dwell time is 0.55 and scrolling frequency is 0.45, after attention mechanism processing, the weights are adjusted to 0.62 and 0.38, reflecting that peak dwell time has a more significant impact on user behavior. Ultimately, the system applies the adjusted weights to the multi-layered behavioral representation, generating a weighted behavioral vector. The calculated comprehensive behavioral vector value for the user is 0.78. This result can be further combined with the preference matching module of the advertising recommendation system to automatically optimize the advertising push strategy. The entire process is automated by the backend system, forming a complete logical chain from data extraction to weight calculation and vector generation, ensuring the accuracy of the analysis and its relevance to business needs.

[0028] Step 104: If the short-term features dominate more than a preset threshold in the obtained weighted behavior vector, the vector balance is adjusted by fusing long-term habit data to obtain an integrated behavior description.

[0029] Step 1: For short-term behavioral features in the weighted behavior vector, determine whether they exceed a preset threshold through proportion analysis. If the proportion of short-term behavior exceeds the preset threshold, obtain long-term habit data from a pre-established behavior database to determine the initial adjustment direction. Step 2: Based on the initial adjustment direction, combine long-term habit data with short-term behavioral features using a data import method to obtain a mixed behavior dataset, yielding an intermediate result for subsequent processing. Step 3: For the mixed behavior dataset in the intermediate result, redistribute the weights of short-term behavior and long-term habits in the behavior vector using a balancing adjustment method to determine the adjusted vector structure. Step 4: Based on the adjusted vector structure, analyze the changes in behavior proportions to obtain updated behavior vectors, resulting in an intermediate description that meets the balance requirements. Step 5: For the updated vectors in the intermediate description, integrate the characteristics of short-term behavior and long-term habits through a comprehensive description construction method to determine the final behavior representation. Step 6: Based on the final behavior representation, analyze the completeness of the adjustment results after vector updates to obtain a comprehensive behavior description.

[0030] Specifically, in processing user behavior data to construct an integrated behavior description, the system analyzes and adjusts the weighted behavior vector using automated information technology to ensure a balance between short-term features and long-term habits. First, the system detects the proportion of short-term features in the weighted behavior vector. For example, if a user's short-term feature, such as a click frequency of 5.2 times per day over the past 7 days, has a proportion of 0.72, exceeding the preset threshold of 0.65, the balance adjustment mechanism is triggered. The system then accesses the long-term habit database, extracting the user's behavior data from the past 90 days, such as an average daily browsing time of 18.4 minutes. Using a weighted average algorithm (0.4 for short-term and 0.6 for long-term), a fusion value is calculated, resulting in an adjusted click frequency impact of 2.1 times / day and a browsing time impact of 11.0 minutes / day. Subsequently, the system uses a vector normalization algorithm to remap the fused short-term and long-term data to the 0-1 range, calculating that the proportion of short-term features decreases to 0.48 and the proportion of long-term features increases to 0.52, forming a new balance vector. Next, the system performs feature smoothing on the integrated vectors, using a sliding window algorithm (with a window size of 3 days) to eliminate short-term fluctuations. Analysis shows that the stability of the adjusted vectors improved to 0.89. Finally, the system stores the integrated behavioral descriptions in the user profile database and interfaces with the content distribution system to automatically optimize personalized recommendation logic. The entire process is completed through a backend automation module, forming a closed-loop processing chain from feature detection to data fusion and vector adjustment, ensuring the comprehensiveness and applicability of the behavioral descriptions.

[0031] Step 105: From the obtained integrated behavior description, obtain the intent change pattern such as interest shift trajectory, determine the user's potential preference for advertising content, and determine the intent vector.

[0032] Step 1: Obtain relevant data on changes in intent from behavioral descriptions, perform preliminary extraction of the change trajectory, and use time series processing to sort out the changes in user patterns and obtain preliminary trajectory records.

[0033] Step 2: Based on the initial trajectory records, analyze the specific direction of interest transfer, split the data according to the transfer pattern, identify the distribution of interest points in different time periods through segmentation processing, and determine the segmentation results of the transfer pattern.

[0034] Step 3: Based on the segmentation results of the transfer mode and the relevant data of the advertising content, analyze the matching of content preferences. Using a comparative processing method, identify the differences in user responses to different advertising content and obtain a preliminary judgment on preference matching.

[0035] Step 4: Based on the initial judgment of preference matching, extract relevant information on potential tendencies. If the intensity of the potential tendency response exceeds the preset threshold, adjust the priority of the tendency judgment through weighted processing to determine the intermediate data of the tendency judgment.

[0036] Step 5: Based on the intermediate data of tendency judgment, combined with the comprehensive information of change trajectory and transfer pattern, construct the basic structure of intention vector using logical mapping to obtain the preliminary representation of intention vector.

[0037] Step Six: Based on the preliminary representation of the intent vector, integrate the data based on the analysis results, refine the structure of the intent vector through multi-dimensional verification, and determine the final form of the intent vector.

[0038] Specifically, in the process of extracting user intent change patterns from integrated behavioral descriptions and judging potential preferences to determine intent vectors, the system gradually completes the analysis and processing through automated information technology. First, the system extracts user interest shift trajectory data from the integrated behavioral descriptions over the past 30 days. Assuming a user's attention to sports ads gradually increases from an initial 0.3 to 0.75 over the past 30 days, while their attention to technology ads decreases from 0.6 to 0.25, the system calculates the interest shift rate using a time series analysis algorithm. It finds that the interest growth rate for sports ads is 0.015 / day, while the decrease rate for technology ads is 0.012 / day, thus determining that the user's interest is shifting positively towards sports content. Subsequently, based on the interest shift trajectory and combined with historical ad interaction data, the system analyzes the user's potential preferences for ad content, extracting click-through rate data for different ad types over the past 60 days. For example, the click-through rate for sports ads is 4.5%, and for technology ads it is 2.1%. Using a Bayesian probability model, the system calculates potential preference scores, resulting in a preference score of 0.82 for sports ads and 0.18 for technology ads, forming a preliminary preference distribution. Next, the system integrates preference distribution and interest shift trajectory data, employing a weighted calculation method (interest shift trajectory weight 0.55, historical click weight 0.45) to derive a comprehensive preference vector. The comprehensive score for sports is 0.78, and for technology, it is 0.22. Finally, the system uses a vector mapping algorithm to transform the comprehensive preference vector into an intent vector, mapping it to the 0-1 range. Combining this with ad inventory data analysis, assuming the current sports ad inventory ratio is 0.65, the system automatically adjusts the weight of sports to 0.7 and the weight of technology to 0.3 in the intent vector, ensuring alignment with actual business resources. The entire process is automated through the backend algorithm module, forming a complete logical chain from interest trajectory extraction to intent vector determination.

[0039] Step 106: Based on the determined intent vector, match the items with the highest similarity in the pre-established ad library to obtain a preliminary recommendation list.

[0040] Step 1: For the initial recommendation set, extract the data of the items with the highest relevance to the intent vector, and use a hierarchical comparison method to sort out the relevance strength between the items and the vector, so as to obtain a priority-ranked list of items.

[0041] Step 2: Based on the priority-sorted list of items, obtain the top-ranked ad resources in the list, use content parsing tools to extract the theme information of the ad resources, and determine the fit between the theme and the intent vector.

[0042] Step 3: For the relevance data between the topic and the intent vector, if the relevance data exceeds the preset threshold, the corresponding ad resource item is retained; if the relevance data is below the preset threshold, the item is removed, resulting in a filtered subset of resources.

[0043] Step 4: Based on the filtered resource subset, obtain the content-related information of each item in the subset, analyze the complementarity between items through multi-angle comparison, and determine the combination of items with high complementarity.

[0044] Step 5: For combinations of items with high complementarity, obtain the distribution of advertising resources within the combination. If the resource distribution within the combination is uneven, adjust the combination structure by re-sorting to obtain an optimized recommended combination.

[0045] Step Six: Based on the optimized recommended combination, obtain the final item data within the combination, and confirm the matching accuracy between the item and the intent vector through logical verification to determine the final recommendation result.

[0046] Specifically, in the process of matching the most similar entries in a pre-established ad library based on the determined intent vector to obtain a preliminary recommendation list, the system completes a series of processes through automated information technology. First, the system reads the determined intent vector; assuming the current user's intent vector is 0.68 for sports and 0.32 for food, the system uses this as baseline data input into the similarity calculation module. Next, the system extracts the feature vectors of ad entries from the pre-established ad library; assuming the feature vector of a sports ad A in the library is 0.72 and the feature vector of a food ad B is 0.35, the system uses a cosine similarity algorithm to calculate the matching degree between the intent vector and the ad entry feature vector. The calculation results show that the similarity between ad A and the user's intent vector is 0.92, and the similarity between ad B and B is 0.88. Subsequently, the system performs batch calculations on all entries in the ad library, filtering out ads with a similarity higher than 0.85 to form a preliminary candidate pool. Assuming the candidate pool contains 10 ads, 7 of which are sports ads and 3 are food ads. To further optimize the recommendation list, the system performs weighted adjustments based on the historical exposure data of each item in the ad library. Assuming sports ad A has a historical exposure rate of 3.2%, while another sports ad C has an exposure rate of 1.8%, the system calculates a comprehensive score using a weighted formula (similarity weighted at 0.6, exposure rate weighted at 0.4). Ad A receives a comprehensive score of 0.75, and ad C receives 0.68. Finally, ads are sorted from highest to lowest comprehensive score, generating a preliminary recommendation list containing the top 5 ads. The entire process is automated through a backend algorithm, ensuring that the recommendation results are highly relevant to user intent while also considering the reasonable allocation of ad resources, forming a complete logical chain from intent vector to recommendation list.

[0047] Step 107: The preliminary recommendation list is dynamically sorted using a sequence model, the priority is adjusted based on real-time user feedback, the final output order is determined, and a personalized advertising sequence is obtained.

[0048] Step 1: Use a sequence model to perform initial sorting on the preliminary recommendation list, obtain the relative position data between items, and determine the preliminary sorting results.

[0049] Step 2: Based on the preliminary ranking results, obtain real-time user feedback data. By analyzing the behavioral preference information in the feedback, if the feedback data indicates that users show high attention to certain items, the priority of the corresponding items will be increased, resulting in an adjusted ranking set.

[0050] Step 3: For the adjusted sorted set, obtain the priority data of each item in the set. By comparing the priorities, if the priority of an item is lower than the preset threshold, move its position to the right to determine the optimized order of the items.

[0051] Step 4: Based on the optimized item order, obtain the data of the top-ranked ad items in the order, use content parsing tools to extract the core information of the items, and determine the set of ad content related to user preferences.

[0052] Step 5: For the set of advertising content, obtain the correlation data between each item in the set. By comparing from multiple perspectives, if some items are found to have overlapping content, remove the duplicate items to obtain the simplified advertising sequence.

[0053] Step Six: Based on the streamlined ad sequence, obtain the distribution of the remaining items in the sequence. Through logical verification, if an item deviates from the theme of the overall sequence, adjust its position to determine the final personalized recommendation order.

[0054] Step 7: For the final personalized recommendation order, obtain the display information of each item in the order, generate a complete advertising presentation plan through data integration tools, and determine the final output advertising sequence.

[0055] Specifically, in the process of dynamically sorting the initial recommendation list and adjusting priorities based on real-time user feedback to generate a personalized ad sequence, the system completes a series of processes through automated algorithms. First, the system obtains an initial recommendation list containing 5 ads. Assuming the initial ranking scores of the ads in the list are 0.82, 0.78, 0.75, 0.71, and 0.68, the system inputs this into a sequence model-based sorting module. A Long Short-Term Memory (LSTM) network algorithm is used to model the ad sequence. Combined with user click behavior data from the past 24 hours, the system analyzes and determines that the user's preference weight for sports ads is 0.65, and their preference weight for entertainment ads is 0.35. Next, the system collects user feedback data in real-time during the current session. Assuming the user's viewing time for the second entertainment ad in the list is 15 seconds, exceeding the average viewing time threshold of 10 seconds, the system, according to the rules of the feedback adjustment module, increases the dynamic priority of this ad by 0.1, from 0.78 to 0.88. Meanwhile, the system detected that the viewing time of the fourth advertisement was only 3 seconds, below the threshold, and its priority was lowered by 0.05, from 0.71 to 0.66. Subsequently, the system combined the sequence model prediction results and the adjusted priority based on real-time feedback to recalculate the final score of the advertisement sequence, resulting in adjusted ranking scores of 0.88, 0.82, 0.75, 0.68, and 0.66, forming a new ranking result. To ensure advertisement diversity, the system also introduced a category balancing mechanism. When it detected that the proportion of sports advertisements was too high, it automatically increased the priority of entertainment advertisements by an additional 0.03, ultimately making the score of the second advertisement 0.91, thus solidifying its first position. The entire process was completed automatically by the backend algorithm, forming a complete logical chain from the initial list to the final personalized sequence, ensuring that the ranking result matches the user's real-time interests while taking into account the rationality of the category distribution.

[0056] This invention provides an intelligent internet advertising promotion system, mainly comprising: The data acquisition module is used to collect data such as browsing time and scrolling speed during the user's advertising interaction process, obtain behavioral sequences to reflect the performance differences in different scenarios, and obtain multi-scale behavioral records. The sequence processing module is used to process continuous interactions in the time dimension using a sequence model based on the acquired behavioral sequences, extract features such as short-term gaze lingering and long-term habits, and determine multi-layered behavioral representations. The feature extraction module is used to apply an attention mechanism to highlight key interaction points, such as the correlation between peak page dwell time and scroll frequency, for a given multi-layer behavior representation, determine the weight distribution between features, and obtain a weighted behavior vector. The attention weighting module is used to adjust the vector balance by fusing long-term habit data if the short-term features dominate the obtained weighted behavior vector more than a preset threshold, so as to obtain an integrated behavior description. The vector integration module is used to extract intent change patterns, such as interest shift trajectories, from the obtained integrated behavior descriptions, determine users' potential preferences for advertising content, and identify intent vectors. The intent analysis module is used to match the most similar items in a pre-built ad library based on the determined intent vector to obtain a preliminary recommendation list; The recommendation ranking module uses a sequence model to dynamically rank the initial recommendation list, adjusts the priority based on real-time user feedback, determines the final output order, and obtains a personalized ad sequence.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent promotion of internet advertising, characterized in that, include, By collecting data such as browsing time and scrolling speed during user interaction with advertisements, behavioral sequences are obtained to reflect performance differences in different scenarios, resulting in multi-scale behavioral records. Based on the acquired behavioral sequences, a sequence model is used to process continuous interactions in the time dimension, extract features such as short-term gaze lingering and long-term habits, and determine multi-layered behavioral representations. For a given multi-layered behavior representation, an attention mechanism is applied to highlight key interaction points such as the correlation between peak page dwell time and scroll frequency, determine the weight distribution among features, and obtain a weighted behavior vector. If the short-term features dominate more than a preset threshold in the obtained weighted behavior vector, the vector balance is adjusted by fusing long-term habit data to obtain an integrated behavior description. From the obtained integrated behavioral descriptions, we can obtain intent change patterns such as interest shift trajectories, determine users' potential preferences for advertising content, and identify intent vectors. Based on the determined intent vector, the most similar items in the pre-built ad library are matched to obtain an initial recommendation list; A sequence model is used to dynamically sort the initial recommendation list, adjust the priority based on real-time user feedback, determine the final output order, and obtain a personalized advertising sequence.

2. The intelligent promotion method for internet advertising according to claim 1, characterized in that, The process involves collecting data such as user browsing time and scrolling speed during ad interaction to obtain behavioral sequences that reflect performance differences in different scenarios, resulting in multi-scale behavioral records, including: By capturing users' browsing time and scrolling speed during ad interactions, initial behavioral sequence data is constructed; Using pre-established data collection tools, relevant information is extracted from interaction records to obtain a preliminary user behavior dataset; Based on the preliminary user behavior dataset, we analyzed the distribution of browsing time and scrolling speed in different scenarios. For duration distribution and speed changes, classification methods are applied to determine the behavioral pattern categories under different scenarios; if the duration distribution of the classified behavioral patterns deviates from the preset threshold in a certain scenario, the data in that scenario is then deeply mined. By further breaking down the performance differences, we can obtain more granular multi-scale data records.

3. The intelligent promotion method for internet advertising according to claim 1, characterized in that, The process involves using a sequence model to process continuous interactions over time based on the acquired behavioral sequence, extracting features such as short-term gaze persistence and long-term habits, and determining multi-layered behavioral representations, including: By obtaining behavioral sequences from user interactions, we can initially break down continuous behaviors over time to obtain short-term and long-term behavioral fragments. Based on short-term and long-term behavioral segments, a sequence model is used to analyze the duration of gaze fixation and identify significant changes in short-term characteristics. For significant changes in short-term characteristics, repetitive patterns in long-term behavioral habits are obtained to determine the stability of long-term characteristics. If the stability of long-term characteristics is lower than a preset threshold, the behavioral segments are regrouped. Based on the regrouped behavioral fragments, key behavioral nodes are extracted from the multi-layered representation to obtain the core behavioral patterns in user interaction. For core behavioral patterns, analyze the distribution of behavior over time to identify potential correlations in continuous behaviors; By identifying potential association points, a multi-layered behavioral representation of user interaction is constructed, resulting in the final set of behavioral features. Based on the set of behavioral features, a behavioral change trajectory over time is generated to determine the behavioral preferences in user interactions.

4. The intelligent promotion method for internet advertising according to claim 1, characterized in that, For the defined multi-layered behavior representation, an attention mechanism is applied to highlight key interaction points, such as the correlation between peak page dwell time and scroll frequency, to determine the weight distribution among features and obtain a weighted behavior vector, including: For key nodes in the behavioral representation, an attention mechanism is used to process the time segments of user operations in layers, identify the correlation pattern between page dwell time and scrolling frequency, and obtain a preliminary distribution of interaction priorities. Based on the initial distribution of interaction priorities, analyze the peak positions of page dwell time, and determine the correlation strength between peak dwell time and frequency by combining the changes in scrolling frequency. To determine the correlation strength between peak dwell times and frequency, feature weights are calculated, and the weight distribution is sorted to identify high-priority interaction areas. By identifying high-priority interactive key areas, the core time segments in user operations are extracted. If the correlation analysis results of the core time segments are lower than the preset threshold, the segments are re-divided to obtain an adjusted set of time segments. Based on the adjusted set of time segments, a weighted vector is constructed, and the behavior representation is updated for the high-value regions in the weight distribution to determine the final weighted behavior pattern. By analyzing the final weighted behavior pattern, the performance of key nodes in different time segments is analyzed to determine the continuity characteristics of user operations and obtain complete interaction analysis results.

5. The intelligent promotion method for internet advertising according to claim 1, characterized in that, If the short-term features dominate more than a preset threshold in the obtained weighted behavior vector, the vector balance is adjusted by fusing long-term habit data to obtain an integrated behavior description, including: For short-term behavioral characteristics in the weighted behavioral vector, the proportion analysis is used to determine whether they exceed the preset threshold. If the proportion of short-term behaviors exceeds a preset threshold, long-term habit data will be obtained from a pre-established behavior database to determine the initial direction of adjustment. Based on the initial adjustment direction, a data introduction approach was adopted to combine long-term habit data with short-term behavioral characteristics to obtain a mixed behavioral dataset, which yielded intermediate results for subsequent processing. For the mixed behavior dataset in the intermediate results, the proportion of short-term behavior and long-term habit in the behavior vector is redistributed through a balancing adjustment method to determine the adjusted vector structure; Based on the adjusted vector structure, analyze the changes in the proportion of behaviors, obtain the updated behavior vectors, and get an intermediate description that meets the balance requirements; For the update vector in the intermediate description, the characteristics of short-term behavior and long-term habits are integrated through a comprehensive description construction method to determine the final behavior representation. Based on the final behavioral representation, the completeness of the results after vector update is analyzed and adjusted to obtain a comprehensive behavioral description.

6. The intelligent promotion method for internet advertising according to claim 1, characterized in that, The process of extracting intent change patterns, such as interest shift trajectories, from the obtained integrated behavioral descriptions, determining users' potential preferences for advertising content, and identifying intent vectors includes: Data related to changes in intent are obtained from behavioral descriptions. The change trajectory is initially extracted, and time series processing is used to sort out the changes in user patterns and obtain preliminary trajectory records. Based on the initial trajectory records, the specific direction of interest shifts is analyzed, and the data is split according to the shift pattern. By segmenting the data, the distribution of interest points in different time periods is identified, and the segmentation results of the shift pattern are determined. Based on the segmentation results of the transfer mode and combined with relevant data of the advertising content, the matching of content preferences is analyzed. By using a comparative processing method, the differences in user responses to different advertising content are identified, and a preliminary judgment on preference matching is obtained. Based on the initial judgment of preference matching, relevant information on potential tendencies is extracted. If the intensity of the potential tendency response exceeds the preset threshold, the priority of the tendency judgment is adjusted through weighted processing to determine the intermediate data of the tendency judgment. Based on the intermediate data of tendency judgment, combined with the comprehensive information of change trajectory and transfer pattern, the basic structure of intention vector is constructed by using logical mapping, and a preliminary representation of intention vector is obtained. Based on the initial representation of the intent vector, the data is integrated according to the analysis results. Through multi-dimensional verification, the structure of the intent vector is refined to determine the final form of the intent vector.

7. The intelligent promotion method for internet advertising according to claim 1, characterized in that, The step involves matching the most similar entries in a pre-established ad library based on the determined intent vector to obtain a preliminary recommendation list, including: For the initial recommendation set, the item data with the highest correlation with the intent vector is obtained. A hierarchical comparison method is used to sort out the correlation strength between the item and the vector, and a priority-ranked list of items is obtained. Based on the priority-sorted list of items, obtain the top-ranked ad resources in the list, use content parsing tools to extract the theme information of the ad resources, and determine the fit data between the theme and the intent vector; If the relevance data between the theme and the intent vector exceeds a preset threshold, the corresponding ad resource entry will be retained. If the matching data is lower than the preset threshold, the entry is removed, and a filtered subset of resources is obtained. Based on the selected resource subset, the content-related information of each item in the subset is obtained. Through multi-angle comparison, the complementarity between items is analyzed to determine the combination of items with high complementarity. For combinations of items with high complementarity, the distribution of advertising resources within the combination is obtained. If the resource distribution within the combination is uneven, the combination structure is adjusted by re-sorting to obtain an optimized recommended combination. Based on the optimized recommended combinations, the final item data within the combinations is obtained. Logical verification is used to confirm the matching accuracy between the items and the intent vector, and the final recommendation result is determined.

8. The intelligent promotion method for internet advertising according to claim 1, characterized in that, The process of dynamically sorting the initial recommendation list using a sequence model, adjusting priorities based on real-time user feedback, determining the final output order, and obtaining a personalized ad sequence includes: A sequence model is used to perform initial sorting on the preliminary recommendation list, and the relative position data between items is obtained to determine the preliminary sorting result. Based on the preliminary ranking results, real-time user feedback data is obtained. By analyzing the behavioral preference information in the feedback, if the feedback data indicates that users show high attention to certain items, the priority of the corresponding items is increased, resulting in an adjusted ranking set. For the adjusted sorted set, obtain the priority data of each item in the set. By comparing the priorities, if the priority of an item is lower than the preset threshold, move its position to the right to determine the optimized order of items. Based on the optimized item order, obtain the data of the top-ranked ad items in the order, use content parsing tools to extract the core information of the items, and determine the set of ad content related to user preferences; For the set of advertising content, obtain the correlation data between each item in the set. Through multi-angle comparison, if some items are found to have overlapping content, remove the duplicate items to obtain a simplified advertising sequence. Based on the streamlined ad sequence, the distribution of the remaining items in the sequence is obtained. Through logical verification, if an item deviates from the theme of the overall sequence, its position is adjusted to determine the final personalized recommendation order. For the final personalized recommendation order, the display information of each item in the order is obtained, and a complete advertising presentation plan is generated through data integration tools to determine the final output advertising sequence.

9. An intelligent internet advertising promotion system, characterized in that, include: The data acquisition module is used to collect data such as browsing time and scrolling speed of users during the advertising interaction process, obtain behavioral sequences to reflect the performance differences in different scenarios, and obtain multi-scale behavioral records. The sequence processing module is used to process continuous interactions in the time dimension using a sequence model based on the acquired behavioral sequences, extract features such as short-term gaze lingering and long-term habits, and determine multi-layered behavioral representations. The feature extraction module is used to apply an attention mechanism to highlight key interaction points, such as the correlation between peak page dwell time and scroll frequency, for a given multi-layer behavior representation, determine the weight distribution between features, and obtain a weighted behavior vector. The attention weighting module is used to adjust the vector balance by fusing long-term habit data if the short-term features dominate the obtained weighted behavior vector more than a preset threshold, so as to obtain an integrated behavior description. The vector integration module is used to extract intent change patterns, such as interest shift trajectories, from the obtained integrated behavior descriptions, determine users' potential preferences for advertising content, and identify intent vectors. The intent analysis module is used to match the most similar items in a pre-built ad library based on the determined intent vector to obtain a preliminary recommendation list; The recommendation ranking module uses a sequence model to dynamically rank the initial recommendation list, adjusts priorities based on real-time user feedback, determines the final output order, and obtains a personalized ad sequence.

Images