Multi-modal data driven intelligent advertisement putting system and method

Abstract

The invention belongs to the technical field of advertisement putting, and discloses a multi-modal data driven intelligent advertisement putting system and method. Comprising the following steps: acquiring and analyzing historical behavior data of a user, extracting a periodic rule of user activity, and identifying candidate delivery windows; when it is detected that the current time is matched with the candidate delivery window, collecting real-time behavior data; performing feature extraction on the real-time behavior data to obtain situation features, and identifying key situation factors to form a user situation portrait; according to the situation portrait of the user, dynamically evaluating interruption cost and response probability, and judging whether the current time is a suitable delivery opportunity, and if so, performing advertisement delivery; according to the invention, comprehensive perception and analysis of user behaviors and situations are realized, and forced pushing under the situation that the user is not suitable for receiving the advertisement is avoided, so that the user experience is improved, and the accuracy and effect of advertisement putting are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of advertising delivery technology, and more specifically, to a multimodal data-driven intelligent advertising delivery system and method. Background Art

[0002] With the continuous development of the Internet and artificial intelligence technologies, intelligent advertising delivery systems based on big data analysis and machine learning have been widely used in multiple scenarios such as social media, e-commerce platforms, and video applications. Existing intelligent advertising delivery systems mainly focus on content matching and audience targeting, that is, based on user interests, behavioral trajectories and content characteristics, they realize intelligent matching and personalized recommendations of advertising materials, thereby improving the relevance and conversion efficiency of advertising delivery. However, in actual applications, the timing of advertising delivery is also a key factor affecting user acceptance and advertising effectiveness. For example, the same advertisement displayed during the user's leisure time and during busy work hours will have significant differences in click-through rate and conversion performance. If the intelligent advertising delivery system lacks the ability to intelligently identify and dynamically adjust the optimal delivery time, even if the advertising content is accurate, it is easy for users to skip or become disgusted due to incorrect delivery timing, thereby affecting the overall delivery effect. Therefore, there is an urgent need for an advertising delivery system with the ability to intelligently judge delivery timing to achieve more efficient and accurate advertising reach and conversion.

[0003] The patent with publication number CN119579258A discloses a digital advertising system; it includes: a data acquisition and preprocessing module, which collects and cleans multi-dimensional user behavior data in real time to ensure high-quality data input; a deep learning model inference module, which builds and updates the context-aware graph, captures the user's immediate needs and complex contextual relationships, and provides an accurate basis for personalized recommendations; a scenario adaptation decision module, which combines real-time contextual information to intelligently judge the user's immediate status and determine the best time to display advertisements; this invention introduces context-aware graph technology into the deep learning model inference module to accurately capture the user's immediate behavior and complex contextual relationships, and efficiently associates the two, thereby providing more personalized and timely advertising content.

[0004] However, although the above-mentioned technology can realize intelligent judgment of the timing of advertising delivery, it is essentially an inference mechanism based on rules and state recognition, and lacks the ability to make refined assessments of the value of delivery. Therefore, there is a risk of forcibly pushing ads even when the user is not suitable to receive them, thereby affecting the accuracy and effectiveness of advertising delivery, leading to problems such as advertising waste, reduced user experience, and lack of flexibility in delivery strategies.

[0005] In view of this, the present invention proposes a multimodal data-driven intelligent advertising delivery system and method to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solution: a multimodal data-driven intelligent advertising delivery method, comprising: Obtain user historical behavior data; Integrating time series pattern mining technology with clustering algorithms, we conduct time series analysis on historical behavior data, extract the periodic patterns of user activity, and identify candidate delivery windows using the Markov decision process. Real-time detection of the matching status between the current time and the candidate delivery window. When it is detected that the current time matches the candidate delivery window, real-time behavior data is collected; Extract features from real-time behavioral data to obtain contextual features, and use the attention mechanism to identify key contextual factors within the contextual features to form a user context profile; Based on the user's contextual profile, the pre-built timing evaluation system dynamically evaluates the interruption cost and response probability, and determines whether the current time is suitable for delivery. If so, the ad is delivered; Collect user response behavior data after receiving advertisements, and dynamically optimize the timing evaluation system based on the response behavior data.

[0007] Furthermore, the historical behavior data includes historical moments collected A set of multimodal behavior data, each set of multimodal behavior data corresponds to a sampling period; multimodal behavior data includes application usage records, page browsing behavior, click behavior, search behavior and device information; The method for extracting user activity periodicity includes: Different digital labels are set for the category data in the historical behavior data and marked as data labels; the pre-trained word embedding model is used to convert the text data in the historical behavior data into corresponding text vectors; each category data in the historical behavior data is replaced with the corresponding data label, and each text data is replaced with the corresponding text vector; the clustering algorithm is used to cluster the historical behavior data to obtain clusters, All multimodal data in each cluster are taken as a data set, and each data set is input into the trained state recognition model to identify the corresponding state label; wherein the state recognition model is a deep neural network model, and the state label is the digital label corresponding to the user state; An active set is preset, which includes active tags corresponding to different user statuses, and the active tags are numerical tags corresponding to the degree of activity; based on the status tags of each data set, the active tags corresponding to each set of multimodal behavior data in the historical behavior data are obtained from the active set; the collection time of each set of multimodal behavior data in the historical behavior data is obtained, and the active tags are arranged in ascending order according to the collection time of the corresponding multimodal behavior data to form an active sequence, and the corresponding collection time is added to each active tag in the active sequence; the active sequence is decomposed by STL, and the active sequence is decomposed into trend terms, seasonal terms and residual terms, and the user activity periodicity is extracted from the seasonal terms; the user activity periodicity includes intraday periodicity, intraweek periodicity and holiday periodicity.

[0008] Furthermore, the method for identifying candidate delivery windows includes: Define a quintuple, which includes the user state space, action space, state transition probability, reward function, and discount factor; The user state space includes different user activity states, including the collection time and active tags. The action space includes advertising and non-advertising. The state transition probability is the probability of a user transitioning to the next active state after receiving an action in the action space while in the current active state. The reward function defines the reward value obtained based on the user's response action after receiving an action in the action space while in the current active state. Response actions include click, close, and ignore. Based on the quintuples, a reinforcement learning algorithm is used to train a reinforcement learning model based on the MDP framework. The trained reinforcement learning model is used to analyze each active tag in the active sequence to generate the optimal strategy corresponding to each active tag. The optimal strategy is the best action selected for each active state in the action space, that is, the optimal strategy includes both advertising and not advertising. The active tag whose optimal strategy is advertising is marked as the delivery tag, and the collection time corresponding to all delivery tags is used as the candidate delivery window.

[0009] Furthermore, the current time is compared with the collection time in the candidate delivery window; if there is a collection time in the candidate delivery window that is the same as the current time, then the current time matches the candidate delivery window; if all collection times in the candidate delivery window are different from the current time, then the current time does not match the candidate delivery window; real-time behavior data is multimodal behavior data collected in real time; The method for obtaining context features includes: The similarity between the real-time behavior data and each cluster is calculated in turn, and each similarity is compared separately. The cluster with the largest similarity is marked as a similar cluster, and the status label corresponding to the similar cluster is used as the real-time status label; according to the real-time status label, the corresponding active label is obtained from the active set and marked as the real-time active label; the application usage records, page browsing behavior, click behavior and search behavior in the real-time behavior data are used as real-time data, and the category data and text data in the real-time data are replaced with corresponding data labels and text vectors respectively. The real-time data is input into the trained attention analysis model to predict the corresponding attention level. The attention analysis model is a deep neural network model; the network connection type, power status and device type in the device information are used as the user environment; based on the real-time active label, attention level, current time and user environment, the contextual features are constructed.

[0010] Furthermore, methods for identifying key situational factors within situational characteristics include: Based on the real-time status label, the self-attention layer of the Transformer architecture is used to calculate the weight coefficient corresponding to each feature in the contextual features in turn. A weight threshold is preset, and each weight coefficient is compared with the weight threshold. Features with a weight coefficient greater than or equal to the weight threshold are considered key contextual factors, while features with a weight coefficient less than the weight threshold are not considered key contextual factors. The method for forming a user context profile includes: Replace the categorical data in the key contextual factors with the corresponding data labels, and combine each key contextual factor and the corresponding weight coefficient into a set of factor vectors, with the factor vectors corresponding to the key contextual factors one-to-one; based on predefined rules, use the rule mapping method to semantically encode each key contextual factor and obtain the semantic label corresponding to each key contextual factor; form a user contextual profile based on the factor vectors, semantic labels, and real-time status labels.

[0011] Further, methods for evaluating the interruption cost and response probability include: A pre-trained word embedding model is used to convert the semantic labels in the user context profile into corresponding semantic vectors. The converted user context profile is input into a trained indicator prediction model to predict the corresponding user response indicator. The indicator prediction model is a deep neural network model. User response indicators include positive response indicators and negative response indicators. Positive response indicators include click-through rate and conversion rate, and negative response indicators include close rate and bounce rate. A fuzzy inference algorithm is used to predict the response probability range based on the positive response indicator, and the interruption cost range based on the negative response indicator. An optimization algorithm is used to search for the first optimal value from the response probability range as the response probability, and the second optimal value from the interruption cost range as the interruption cost. Methods for determining whether the current time is suitable for delivery include: A preset ratio set includes a proportional coefficient corresponding to the interruption cost and the response probability; the response probability and the interruption cost are multiplied by the corresponding proportional coefficient in the ratio set respectively to obtain the response ratio and the interruption ratio, and the difference between the response ratio and the interruption ratio is used as the response benefit ratio; obtain Historical earnings ratio, The historical return ratio is the most recent one to the current time The response-return ratio corresponding to the historical appropriate delivery time, is an integer greater than 1; the preset factor set includes weighting factors, The weight factors increase in sequence; obtain the historical time corresponding to each historical return ratio, and then The historical return ratio and the factor set weight factors to correspond; Each historical return ratio is multiplied by the corresponding weight factor in the factor set, and the average is processed to obtain the return ratio threshold; the response return ratio is compared with the return ratio threshold. If the response return ratio is greater than or equal to the return ratio threshold, the current time is regarded as the appropriate time for delivery. If the response return ratio is less than the return ratio threshold, the current time is not regarded as the appropriate time for delivery.

[0012] Furthermore, the step of searching for a first optimal value from the response probability range as the response probability includes: Step S101: Preset the division coefficient , ; Step S102: Based on the division coefficient , divide the response probability range into two split points and ; Step S103: Calculate two segmentation points respectively and The corresponding matching coefficient will split the points The corresponding matching coefficient is marked as the first coefficient, and the split point The corresponding matching coefficient is marked as the second coefficient; Step S104: Compare the first coefficient with the second coefficient; if the first coefficient is greater than or equal to the second coefficient, proceed to step S105; if the first coefficient is less than the second coefficient, proceed to step S106; Step S105: Calculate the matching coefficient corresponding to the maximum value of the response probability range and mark it as the third coefficient; update the maximum value of the response probability range as the split point , the minimum value of the response probability range remains unchanged; subtract the third coefficient from the second coefficient to obtain the coefficient difference, and proceed to step S107; Step S106: Calculate the matching coefficient corresponding to the minimum value of the response probability range and mark it as the fourth coefficient, and update the minimum value of the response probability range as the split point , the maximum value of the response probability range remains unchanged; subtract the fourth coefficient from the first coefficient to obtain the coefficient difference, and proceed to step S107; Step S107: Determine whether the segmentation is completed. If the segmentation is completed, proceed to step S108. If the segmentation is not completed, update the segmentation coefficient according to the coefficient difference, and re-divide two segmentation points from the response probability range according to the updated segmentation coefficient. and , and returns to step S103; Step S108: Obtain a first optimal value based on the average of the maximum value and the minimum value corresponding to the response probability range, and use it as the response probability; The step of searching for the second optimal value from the interruption cost range as the interruption cost is consistent with the step of searching for the first optimal value from the response probability range as the response probability.

[0013] Furthermore, in step S102, the segmentation point The expression is: ; Split point The expression is: ;in, is the minimum value of the response probability range, is the maximum value of the response probability range; In step S107, the method for determining whether the segmentation is completed includes: A width threshold is preset, and the width threshold is a real number greater than 0; the maximum value of the updated response probability range is subtracted from the corresponding minimum value to obtain the interval width; the interval width is compared with the width threshold; if the interval width is greater than or equal to the width threshold, the segmentation is not completed; if the interval width is less than the width threshold, the segmentation is completed; Methods for updating the split coefficients based on coefficient differences include: Preset adjustment coefficient and adjustment threshold, the adjustment coefficient is The adjustment threshold is an integer greater than 0; the coefficient difference is compared with the adjustment threshold; if the coefficient difference is greater than or equal to the adjustment threshold, the coefficient difference is multiplied by the adjustment coefficient, and the division coefficient is added to obtain a first adjustment coefficient, and the division coefficient is updated according to the first adjustment coefficient; if the coefficient difference is less than the adjustment threshold, the coefficient difference is multiplied by the adjustment coefficient to obtain an adjustment component, the division coefficient is subtracted from the adjustment component to obtain a second adjustment coefficient, and the division coefficient is updated according to the second adjustment coefficient.

[0014] Furthermore, the response behavior data includes click behavior, conversion behavior, closing behavior and bounce behavior; When the response behavior data is click behavior or conversion behavior, the timing evaluation system will not be dynamically optimized; When the response behavior data is a close behavior or a bounce behavior, the methods for dynamically optimizing the timing evaluation system include: Preset the optimization factor, add the product of the split coefficient and the optimization factor to the split coefficient to obtain the optimization coefficient; subtract the product of the width threshold and the optimization factor from the width threshold to obtain the optimization threshold; obtain the coefficient optimization range based on the split coefficient and the optimization coefficient, that is, the split coefficient is the minimum value of the coefficient optimization range, and the optimization coefficient is the maximum value of the coefficient optimization range; obtain the threshold optimization range based on the width threshold and the optimization threshold, that is, the width threshold is the maximum value of the threshold optimization range, and the optimization threshold is the minimum value of the threshold optimization range; randomly select a value from each of the coefficient optimization range and the threshold optimization range, and use them as a set of optimization sets to obtain a total of Group optimization collection, is an integer greater than 1; According to each group of optimization sets, the response probability and interruption cost corresponding to each group of optimization sets are recalculated; the matching coefficient corresponding to the response probability of each group of optimization sets is calculated and marked as the forward coefficient; the matching coefficient corresponding to the interruption cost of each group of optimization sets is calculated and marked as the reverse coefficient; the forward coefficient and reverse coefficient of each group of optimization sets are averaged to obtain the coefficient mean; all coefficient means are compared separately, and the optimization set with the largest coefficient mean is marked as the optimal set; the segmentation coefficient and width threshold are dynamically optimized according to the values ​​in the optimal set.

[0015] A multimodal data-driven intelligent advertising delivery system implements the multimodal data-driven intelligent advertising delivery method, comprising: Data acquisition module, used to obtain users' historical behavior data; The activity analysis module integrates time series pattern mining technology and clustering algorithms to perform time series analysis on historical behavior data, extract the periodic patterns of user activity, and identify candidate delivery windows using the Markov decision process. The data collection module is used to detect the matching status of the current time and the candidate delivery window in real time. When it is detected that the current time matches the candidate delivery window, real-time behavior data is collected; The context building module is used to extract features from real-time behavioral data, obtain context features, and use the attention mechanism to identify key context factors within the context features to form a user context profile; The timing decision module is used to dynamically evaluate the interruption cost and response probability based on the user's contextual profile through a pre-built timing evaluation system, and to determine whether the current time is suitable for delivery. If so, the ad is delivered; The feedback optimization module is used to collect the user's response behavior data after receiving the advertisement, and dynamically optimize the timing evaluation system based on the response behavior data.

[0016] The technical effects and advantages of the multimodal data-driven intelligent advertising delivery system and method of the present invention are as follows: By integrating time series analysis of historical behavioral data with user status identification, we extract the periodic patterns of user activity, and dynamically identify accurate candidate delivery windows in combination with the Markov decision process, thereby improving the accuracy of timing selection for advertising reach. We use the attention mechanism to analyze real-time behavioral data, accurately construct a portrait of the user's current situation, and intelligently determine the best time for advertising delivery through dynamic evaluation of interruption costs and response probabilities, thereby improving the effectiveness of advertising delivery. Through continuous tracking of response behavior data and dynamic optimization of the timing evaluation system, we achieve intelligent adjustment and continuous improvement of the advertising delivery process, further enhancing the relevance and conversion rate of advertising delivery. We achieve comprehensive perception and analysis of user behavior and situations, fundamentally changing the limitations of traditional advertising delivery that relies too much on content matching and audience selection, avoiding forced push in situations where users are not suitable to receive ads, thereby improving user experience and effectively improving the accuracy and effectiveness of advertising delivery. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic diagram of a multimodal data-driven intelligent advertising delivery system according to Example 1 of the present invention; Figure 2 This is a flow chart of a multimodal data-driven intelligent advertising delivery method according to embodiment 2 of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1 See also Figure 1As shown, this embodiment is a multimodal data-driven intelligent advertising delivery system, including a data acquisition module, an activity analysis module, a data collection module, a scenario construction module, a timing decision module and a feedback optimization module; each module is connected by wired and / or wireless means to realize data transmission between modules.

[0020] The data acquisition module is used to obtain the user's historical behavior data.

[0021] Historical behavior data includes historical moments A set of multimodal behavior data, which is a record of the behavior generated during the interaction between users and the platform. is an integer greater than 1; each set of multimodal behavior data corresponds to a sampling period, that is, a set of multimodal behavior data is collected every time a sampling period passes. The sampling period is pre-set by those skilled in the art according to actual conditions, such as ten minutes, twenty minutes, etc.; multimodal behavior data includes application usage records, page browsing behavior, click behavior, search behavior, device information, etc.; application usage records include usage time, usage frequency, etc.; page browsing behavior includes page dwell time, page content, page scroll depth (that is, the distance the user scrolls down when browsing the page, which is used to measure the user's interest in the page content). click behavior includes ad click records, link click records, etc.; search behavior includes search keywords, search frequency, search time, etc.; device information includes network connection type, battery status, device type, operating system, etc.; historical behavior data is obtained through local logs (that is, application usage records or system log files saved locally on the user's device, usually automatically generated by the operating system or the application itself), front-end embedding (that is, inserting collection code in the front-end code to obtain multimodal behavior data), etc.

[0022] It should be noted that application usage records are used to reflect the time patterns of users' daily use of applications, including high-frequency usage periods and continuously active time windows; page browsing behavior is used to show the periodic patterns of users' content consumption, indicating regular periods of user concentration and deep participation; click behavior is used to reveal the temporal distribution characteristics of users' willingness to interact, and identify periodic moments when users tend to respond more readily; search behavior is used to indicate the regular rhythm of users' active information acquisition, reflecting the periodic changes in demand; device information is used to indicate the periodic changes in the user's scenario, revealing regular activities in daily life and work through device usage patterns.

[0023] The activity analysis module is used to integrate time series pattern mining technology and clustering algorithms to perform time series analysis on historical behavior data, extract the periodic patterns of user activity, and identify candidate delivery windows in combination with the Markov decision process.

[0024] Methods for extracting periodic patterns of user activity include: Different digital labels are set for the category data (such as network connection type, device type, operating system, etc.) in the historical behavior data, and marked as data labels; pre-trained word embedding models (such as Word2Vec, GloVe, FastText, etc.) are used to convert the text data (such as page content, search keywords, etc.) in the historical behavior data into corresponding text vectors; each category data in the historical behavior data is replaced with the corresponding data label, and each text data is replaced with the corresponding text vector; clustering algorithms (such as K-means clustering algorithm, hierarchical clustering algorithm, density-based spatial clustering algorithm, etc.) are used to cluster the historical behavior data to obtain clusters, , each cluster contains at least one set of multimodal behavior data; all multimodal data in each cluster are taken as a set of data sets, and the data sets correspond to the clusters one by one; each set of data sets is input into the trained state recognition model to identify the corresponding state label; wherein the state recognition model is a deep neural network model, and the deep neural network model is an existing technology, and the specific training process is not described in detail here; the state label is a digital label corresponding to the user state, and the state labels of different user states are different; user states include focused work state (i.e., long page dwell time, deep page scrolling depth, high search frequency, WiFi connection, etc.), casual browsing state (i.e., short page dwell time, clicking on different types of ads or links, low search frequency, etc.), quick query state (i.e., extremely short page dwell time, high page bounce rate, short usage time, any network type, etc.), etc.; Preset active set, which includes active tags corresponding to different user status. Active tags are digital tags corresponding to activity levels, and different activity levels correspond to different digital tags. Activity levels include high activity, medium activity, low activity, etc.; according to the status tags of each data set, obtain the active tags corresponding to each group of multimodal behavior data in the historical behavior data from the active set; obtain the collection time of each group of multimodal behavior data in the historical behavior data, obtain the timestamp of the collection time through local logs or front-end embedding, and obtain it through JavaScript conversion. The collection time is, for example, 9 am on Mondays on weekdays and 8 pm on Saturdays on holidays; the active tags are based on the collection time of the corresponding multimodal behavior data Arrange the data in ascending order (i.e., active tags with earlier collection times are ranked earlier, and active tags with later collection times are ranked later) to form an active sequence, and add the corresponding collection time to each active tag in the active sequence; perform STL decomposition on the active sequence to decompose the active sequence into trend terms, seasonal terms, and residual terms, and extract the user activity periodicity from the seasonal terms; user activity periodicity includes intraday periodicity (i.e., user activity fluctuations within a day), intraweek periodicity (i.e., user activity fluctuations within a week), and holiday periodicity (i.e., user activity fluctuations during holidays and weekdays); it should be noted that STL decomposition is an existing technology, and the specific process will not be elaborated on here.

[0025] Methods for identifying candidate delivery windows include: Define a quintuple, which includes the user state space, action space, state transition probability, reward function, and discount factor; The user state space includes different activity states of the user, which include the collection time and activity label, such as medium activity at 9:00 a.m. on Monday morning and high activity at 8:00 p.m. on Saturday, a holiday. The action space includes advertising and no advertising. The state transition probability is the probability that a user will transition to the next activity state after receiving an action in the action space in the current activity state. For example, after receiving an ad at medium activity at 9:00 a.m. on Monday morning, the probability of transitioning to high activity at 10:00 a.m. on Monday morning is 0.7. The state transition probability is obtained by analyzing and calculating historical behavior data by those skilled in the art. The reward function defines the reward value obtained based on the user's response action after receiving an action in the action space in the current activity state. Response actions include click, close, and ignore. The reward function is set by those skilled in the art based on historical behavior data. For example, the reward value for click is +5 points, for close is -3 points, and for ignore is -1 point. If no ad is served, the reward value is 0. The discount factor is used to represent the trade-off between current rewards and future rewards. A larger value indicates a greater emphasis on long-term benefits. It is pre-set by those skilled in the art based on actual conditions and is usually between 0.8 and 0.99. Based on the quintuples, a reinforcement learning algorithm (such as Q-learning, deep Q-network DQN, etc.) is used to train a reinforcement learning model based on the MDP framework. The trained reinforcement learning model is used to analyze each active tag in the active sequence to generate the optimal strategy corresponding to each active tag. The optimal strategy is the optimal action selected for each active state in the action space, that is, the optimal strategy includes both advertising and not advertising. The active tags whose optimal strategy is advertising are marked as delivery tags, and the collection times corresponding to all delivery tags are used as candidate delivery windows. It should be noted that the reinforcement learning model based on the MDP framework is an existing technology, and the specific training process will not be described in detail here.

[0026] The data collection module is used to detect the matching status of the current time and the candidate delivery window in real time. When it is detected that the current time matches the candidate delivery window, real-time behavior data is collected.

[0027] Compare the current time with the collection time in the candidate delivery window; if there is a collection time in the candidate delivery window that is the same as the current time, then the current time matches the candidate delivery window; if all collection times in the candidate delivery window are different from the current time, then the current time does not match the candidate delivery window; Real-time behavior data is multimodal behavior data collected in real time, corresponding to the most recent sampling period.

[0028] The context construction module is used to extract features from real-time behavioral data, obtain context features, and use the attention mechanism to identify key context factors within the context features to form a user context profile.

[0029] Methods for obtaining context features include: The similarity between the real-time behavior data and each cluster is calculated in turn, and each similarity is compared separately. The cluster with the largest similarity is marked as a similar cluster, and the status label corresponding to the similar cluster is used as the real-time status label. The similarity calculation method is the existing technology, and the specific calculation process is not described in detail here. According to the real-time status label, the corresponding active label is obtained from the active set and marked as the real-time active label. The application usage records, page browsing behavior, click behavior and search behavior in the real-time behavior data are used as real-time data, and the category data and text data in the real-time data are replaced with corresponding data labels and text vectors respectively. The real-time data is input into the trained attention analysis model to predict the corresponding attention level. The attention analysis model is a deep neural network model. The network connection type, power status and device type in the device information are used as the user environment. According to the real-time active label, attention level, current time and user environment, the context feature is constructed.

[0030] Methods for identifying key situational factors within situational characteristics include: According to the real-time status label, the self-attention layer of the Transformer architecture is used to calculate the weight coefficient corresponding to each feature in the contextual features in turn; the self-attention layer of the Transformer architecture is a prior art, and the specific calculation process will not be described in detail here; a weight threshold is preset, and the weight threshold is pre-set by technical personnel in this field according to actual conditions; each weight coefficient is compared with the weight threshold respectively, and the features with weight coefficients greater than or equal to the weight threshold are regarded as key contextual factors, and the features with weight coefficients less than the weight threshold are not regarded as key contextual factors; it should be understood that under different user states, the key factors affecting user advertising acceptance are different, so it is necessary to dynamically identify the key contextual factors in the contextual features, so as to improve the accuracy and efficiency of advertising delivery decisions.

[0031] Methods for forming user context profiles include: The categorical data in the key contextual factors are replaced with corresponding data labels, and each key contextual factor and the corresponding weight coefficient are combined into a set of factor vectors, and the factor vectors correspond to the key contextual factors one-to-one; based on predefined rules, a rule mapping method is used to semantically encode each key contextual factor to obtain the semantic label corresponding to each key contextual factor; for example, the semantic label for a high level of attention is a focused state, and the semantic label for a WiFi connection is a stable network environment, etc.; based on the factor vectors, semantic labels and real-time status labels, a user contextual portrait is formed; it should be noted that the predefined rules used in the rule mapping method are pre-defined by those skilled in the art according to actual conditions.

[0032] The timing decision module is used to dynamically evaluate the interruption cost and response probability based on the user's situational profile through a pre-built timing evaluation system, and to determine whether the current time is suitable for delivery. If it is a suitable time for delivery, the advertisement will be delivered.

[0033] Interruption cost is the degree to which advertising interferes with the user's current behavior. Response probability is the probability that the user will actively respond to the ad, such as clicking on the ad or staying to watch. Methods for assessing the cost of disruption and the probability of response include: A pre-trained word embedding model is used to convert the semantic labels in the user context profile into corresponding semantic vectors; the converted user context profile is input into the trained indicator prediction model to predict the corresponding user response indicator, which is a deep neural network model; user response indicators include positive response indicators and negative response indicators, positive response indicators include click-through rate and conversion rate, and negative response indicators include close rate and bounce rate; the click-through rate is the probability that a user clicks on an ad after the ad is displayed, the conversion rate is the probability that a user completes the target behavior (such as purchase, registration, download, etc.) after clicking on the ad, the close rate is the probability that a user closes the ad after the ad is displayed, and the bounce rate is the probability that a user does not interact with the ad (such as click, close, etc.) after the ad is displayed; a fuzzy inference algorithm is used to predict the response probability range based on the positive response indicator and the interruption cost range based on the negative response indicator; an optimization algorithm is used to search for the first optimal value from the response probability range as the response probability, and search for the second optimal value from the interruption cost range as the interruption cost; the fuzzy inference algorithm is a prior art, and the specific process will not be described in detail here.

[0034] The step of searching for a first optimal value from the response probability range as the response probability includes: Step S101: Preset the division coefficient , ; Step S102: Based on the division coefficient , divide the response probability range into two split points and ; Step S103: Calculate two segmentation points respectively and The corresponding matching coefficient will split the points The corresponding matching coefficient is marked as the first coefficient, and the split point The corresponding matching coefficient is marked as the second coefficient; Step S104: Compare the first coefficient with the second coefficient; if the first coefficient is greater than or equal to the second coefficient, proceed to step S105; if the first coefficient is less than the second coefficient, proceed to step S106; Step S105: Calculate the matching coefficient corresponding to the maximum value of the response probability range and mark it as the third coefficient; update the maximum value of the response probability range as the split point , the minimum value of the response probability range remains unchanged; subtract the third coefficient from the second coefficient to obtain the coefficient difference, and proceed to step S107; Step S106: Calculate the matching coefficient corresponding to the minimum value of the response probability range and mark it as the fourth coefficient, and update the minimum value of the response probability range as the split point , the maximum value of the response probability range remains unchanged; subtract the fourth coefficient from the first coefficient to obtain the coefficient difference, and proceed to step S107; Step S107: Determine whether the segmentation is completed. If the segmentation is completed, proceed to step S108. If the segmentation is not completed, update the segmentation coefficient according to the coefficient difference, and re-divide two segmentation points from the response probability range according to the updated segmentation coefficient. and , and returns to step S103; Step S108: Obtain a first optimal value according to the average of the maximum value and the minimum value corresponding to the response probability range, and use it as the response probability.

[0035] The step of searching for the second optimal value from the interruption cost range as the interruption cost is consistent with the step of searching for the first optimal value from the response probability range as the response probability.

[0036] In the above step S102, the segmentation point The expression is: ; Split point The expression is: ;in, is the minimum value of the response probability range, is the maximum value of the response probability range.

[0037] The method for calculating the matching coefficient in the above step S103 is: using the numerical value of the segmentation point and the positive response index as calculation data, inputting the calculation data into a trained coefficient prediction model, and predicting the corresponding matching coefficient. The matching coefficient is the degree of matching between the numerical value of the segmentation point and the positive response index. The coefficient prediction model is a deep neural network model. In step S107, the method for determining whether the segmentation is completed includes: A preset width threshold is preset by a person skilled in the art based on actual conditions, and the width threshold is a real number greater than 0; the maximum value of the updated response probability range is subtracted from the corresponding minimum value to obtain the interval width; the interval width is compared with the width threshold; if the interval width is greater than or equal to the width threshold, the segmentation is not completed; if the interval width is less than the width threshold, the segmentation is completed.

[0038] Methods for updating the split coefficients based on coefficient differences include: The preset adjustment coefficient and adjustment threshold are preset by those skilled in the art according to the actual situation. The adjustment coefficient is The adjustment threshold is an integer greater than 0; the coefficient difference is compared with the adjustment threshold; if the coefficient difference is greater than or equal to the adjustment threshold, the coefficient difference is multiplied by the adjustment coefficient, and the division coefficient is added to obtain a first adjustment coefficient, and the division coefficient is updated according to the first adjustment coefficient; if the coefficient difference is less than the adjustment threshold, the coefficient difference is multiplied by the adjustment coefficient to obtain an adjustment component, the division coefficient is subtracted from the adjustment component to obtain a second adjustment coefficient, and the division coefficient is updated according to the second adjustment coefficient.

[0039] Methods for determining whether the current time is suitable for delivery include: A preset ratio set includes a proportional coefficient corresponding to the interruption cost and the response probability. The ratio set is preset by those skilled in the art according to actual conditions. The response probability and the interruption cost are multiplied by the corresponding proportional coefficient in the ratio set respectively to obtain the response ratio and the interruption ratio, and the difference between the response ratio and the interruption ratio is used as the response benefit ratio. Historical earnings ratio, The historical return ratio is the most recent one to the current time The response-return ratio corresponding to the historical appropriate delivery time, is an integer greater than 1; Preset factor set, the factor set includes weighting factors, The weight factors increase in sequence, and the factor set is pre-set by those skilled in the art according to the actual situation; obtain the historical time corresponding to each historical return ratio, and The historical return ratio and the factor set The historical time is the time when the historical return ratio is obtained. The closer the historical time is to the current time, the larger the weight factor corresponding to the historical return ratio. That is, the weight factor corresponding to the historical return ratio of the historical time closest to the current time is the largest, and the weight factor corresponding to the historical return ratio of the historical time farthest from the current time is the smallest. It should be noted that The historical revenue ratios and corresponding historical times are obtained from the database built into the intelligent advertising delivery system.

[0040] Will The historical return ratios are multiplied by the corresponding weight factors in the factor set, and the average is processed to obtain the return ratio threshold; that is, The historical return ratios are multiplied by the corresponding weight factors in the factor set, added up in sequence, and then divided by , obtain the benefit ratio threshold; compare the response benefit ratio with the benefit ratio threshold. If the response benefit ratio is greater than or equal to the benefit ratio threshold, the current time is regarded as a suitable time for delivery; if the response benefit ratio is less than the benefit ratio threshold, the current time is not regarded as a suitable time for delivery.

[0041] The feedback optimization module is used to collect the user's response behavior data after receiving the advertisement, and dynamically optimize the timing evaluation system based on the response behavior data.

[0042] Response behavior data refers to the interactive or non-interactive behavior data generated by users after an ad is displayed, reflecting the user's acceptance of the ad content, interest intention, and subsequent action tendency. Response behavior data includes click behavior (i.e., users click on the ad), conversion behavior (i.e., users purchase, register, download, etc. on the ad page), close behavior (i.e., users close the ad), and bounce behavior (i.e., users do not interact with the ad). When the response behavior data is click behavior or conversion behavior, the timing evaluation system will not be dynamically optimized; When the response behavior data is a close behavior or a bounce behavior, it indicates that the user lacks interest in the ad. Therefore, the methods for dynamically optimizing the timing evaluation system include: The optimization factor is preset and is pre-set by those skilled in the art according to actual conditions; the product of the split coefficient and the optimization factor is added to the split coefficient to obtain the optimization coefficient; the product of the width threshold and the optimization factor is subtracted from the width threshold to obtain the optimization threshold; according to the split coefficient and the optimization coefficient, the coefficient optimization range is obtained, that is, the split coefficient is the minimum value of the coefficient optimization range, and the optimization coefficient is the maximum value of the coefficient optimization range; according to the width threshold and the optimization threshold, the threshold optimization range is obtained, that is, the width threshold is the maximum value of the threshold optimization range, and the optimization threshold is the minimum value of the threshold optimization range; a value is randomly selected from each of the coefficient optimization range and the threshold optimization range, and a total of Group optimization collection, is an integer greater than 1; According to each group of optimization sets, the response probability and interruption cost corresponding to each group of optimization sets are recalculated; the matching coefficient corresponding to the response probability of each group of optimization sets is calculated and marked as the forward coefficient; the matching coefficient corresponding to the interruption cost of each group of optimization sets is calculated and marked as the reverse coefficient; the forward coefficient and reverse coefficient of each group of optimization sets are averaged to obtain the coefficient mean; all coefficient means are compared separately, and the optimization set with the largest coefficient mean is marked as the optimal set; the segmentation coefficient and width threshold are dynamically optimized according to the values ​​in the optimal set.

[0043] It should be understood that since the split coefficient is in the interval Inside The width threshold is a real number greater than 0, so the optimization coefficient is also in the interval The optimization threshold is a real number greater than 0.

[0044] This embodiment extracts the periodic patterns of user activity by integrating time series analysis of historical behavioral data and user status identification, and dynamically identifies accurate candidate delivery windows in combination with the Markov decision process, thereby improving the accuracy of timing selection for advertising reach; it uses an attention mechanism to analyze real-time behavioral data, accurately constructs a portrait of the user's current situation, and intelligently determines the best time for advertising delivery by dynamically evaluating interruption costs and response probabilities, thereby improving the effectiveness of advertising delivery; through continuous tracking of response behavior data and dynamic optimization of the timing evaluation system, it achieves intelligent adjustment and continuous improvement of the advertising delivery process, further enhancing the relevance and conversion rate of advertising delivery; it achieves comprehensive perception and analysis of user behavior and context, fundamentally changing the limitations of traditional advertising delivery that relies too much on content matching and audience selection, avoiding forced push in situations where users are not suitable to receive advertisements, thereby improving user experience and effectively improving the accuracy and effectiveness of advertising delivery.

[0045] Example 2 See also Figure 2 As shown, for the parts not described in detail in this embodiment, please refer to the description of Example 1. A multimodal data-driven intelligent advertising delivery method is provided, and the method includes: Obtain user historical behavior data; Integrating time series pattern mining technology with clustering algorithms, we conduct time series analysis on historical behavior data, extract the periodic patterns of user activity, and identify candidate delivery windows using the Markov decision process. Real-time detection of the matching status between the current time and the candidate delivery window. When it is detected that the current time matches the candidate delivery window, real-time behavior data is collected; Extract features from real-time behavioral data to obtain contextual features, and use the attention mechanism to identify key contextual factors within the contextual features to form a user context profile; Based on the user's contextual profile, the pre-built timing evaluation system dynamically evaluates the interruption cost and response probability, and determines whether the current time is suitable for delivery. If so, the ad is delivered; Collect user response behavior data after receiving advertisements, and dynamically optimize the timing evaluation system based on the response behavior data.

[0046] Example 3 The present application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories may store computer-readable code that, when executed by the one or more processors, may execute the multimodal data-driven intelligent advertising delivery method described above.

[0047] The method or system according to the embodiment of the present application can also be implemented with the help of the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or a hard disk, can store a multimodal data-driven intelligent advertising delivery method provided by this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is only exemplary. When implementing different devices, one or more components in the electronic device shown in this application can be omitted according to actual needs.

[0048] Example 4 One embodiment of the present application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, a multimodal data-driven intelligent advertising delivery method according to an embodiment of the present application described with reference to the above figures can be executed. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), a hard disk, flash memory, etc.

[0049] In addition, according to embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the present application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to execute instructions corresponding to the steps of the method provided in the present application, such as a multimodal data-driven intelligent advertising delivery method. When the computer program is executed by a central processing unit (CPU), the above-mentioned functions defined in the method of the present application are performed.

[0050] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will be able to modify the technical solutions described in the foregoing embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0051] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0052] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0053] In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0054] In the description of the present invention, “several” means one or more, and “a large number” means two or more.

[0055] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0056] The formulas in this manual are all dimensionless and calculated using numerical values. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field based on actual conditions.

[0057] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A multimodal data-driven intelligent advertising delivery method, characterized in that: include: Obtain user historical behavior data; Integrating time series pattern mining technology with clustering algorithms, we conduct time series analysis on historical behavior data, extract the periodic patterns of user activity, and identify candidate delivery windows using the Markov decision process. Real-time detection of the matching status between the current time and the candidate delivery window. When it is detected that the current time matches the candidate delivery window, real-time behavior data is collected; Extract features from real-time behavioral data to obtain contextual features, and use the attention mechanism to identify key contextual factors within the contextual features to form a user context profile; Based on the user's contextual profile, the pre-built timing evaluation system dynamically evaluates the interruption cost and response probability, and determines whether the current time is suitable for delivery. If so, the ad is delivered; Collect user response behavior data after receiving advertisements, and dynamically optimize the timing evaluation system based on the response behavior data.

2. The multimodal data-driven intelligent advertising delivery method according to claim 1, characterized in that: The historical behavior data includes the historical moments collected A set of multimodal behavior data, each set of multimodal behavior data corresponds to a sampling period; multimodal behavior data includes application usage records, page browsing behavior, click behavior, search behavior and device information; The method for extracting user activity periodicity includes: Different digital labels are set for the category data in the historical behavior data and marked as data labels; the pre-trained word embedding model is used to convert the text data in the historical behavior data into corresponding text vectors; each category data in the historical behavior data is replaced with the corresponding data label, and each text data is replaced with the corresponding text vector; the clustering algorithm is used to cluster the historical behavior data to obtain clusters, All multimodal data in each cluster are taken as a data set, and each data set is input into the trained state recognition model to identify the corresponding state label; wherein the state recognition model is a deep neural network model, and the state label is the digital label corresponding to the user state; An active set is preset, which includes active tags corresponding to different user statuses, and the active tags are numerical tags corresponding to the degree of activity; based on the status tags of each data set, the active tags corresponding to each set of multimodal behavior data in the historical behavior data are obtained from the active set; the collection time of each set of multimodal behavior data in the historical behavior data is obtained, and the active tags are arranged in ascending order according to the collection time of the corresponding multimodal behavior data to form an active sequence, and the corresponding collection time is added to each active tag in the active sequence; the active sequence is decomposed by STL, and the active sequence is decomposed into trend terms, seasonal terms and residual terms, and the user activity periodicity is extracted from the seasonal terms; the user activity periodicity includes intraday periodicity, intraweek periodicity and holiday periodicity.

3. The multimodal data-driven intelligent advertising delivery method according to claim 2, characterized in that: Methods for identifying candidate delivery windows include: Define a quintuple, which includes the user state space, action space, state transition probability, reward function, and discount factor; The user state space includes different user activity states, including the collection time and active tags. The action space includes advertising and non-advertising. The state transition probability is the probability of a user transitioning to the next active state after receiving an action in the action space while in the current active state. The reward function defines the reward value obtained based on the user's response action after receiving an action in the action space while in the current active state. Response actions include click, close, and ignore. Based on the quintuples, a reinforcement learning algorithm is used to train a reinforcement learning model based on the MDP framework. The trained reinforcement learning model is used to analyze each active tag in the active sequence to generate the optimal strategy corresponding to each active tag. The optimal strategy is the best action selected for each active state in the action space, that is, the optimal strategy includes both advertising and not advertising. The active tag whose optimal strategy is advertising is marked as the delivery tag, and the collection time corresponding to all delivery tags is used as the candidate delivery window.

4. The multimodal data-driven intelligent advertising delivery method according to claim 3, characterized in that: Compare the current time with the collection time in the candidate delivery window; if there is a collection time in the candidate delivery window that is the same as the current time, then the current time matches the candidate delivery window; if all collection times in the candidate delivery window are different from the current time, then the current time does not match the candidate delivery window; Real-time behavioral data is multimodal behavioral data collected in real time; The method for obtaining context features includes: The similarity between the real-time behavior data and each cluster is calculated in turn, and each similarity is compared separately. The cluster with the largest similarity is marked as a similar cluster, and the status label corresponding to the similar cluster is used as the real-time status label; according to the real-time status label, the corresponding active label is obtained from the active set and marked as the real-time active label; the application usage records, page browsing behavior, click behavior and search behavior in the real-time behavior data are used as real-time data, and the category data and text data in the real-time data are replaced with corresponding data labels and text vectors respectively. The real-time data is input into the trained attention analysis model to predict the corresponding attention level. The attention analysis model is a deep neural network model; the network connection type, power status and device type in the device information are used as the user environment; based on the real-time active label, attention level, current time and user environment, the contextual features are constructed.

5. The multimodal data driven intelligent advertising delivery method according to claim 4, characterized in that: Methods for identifying key situational factors within situational characteristics include: Based on the real-time status label, the self-attention layer of the Transformer architecture is used to calculate the weight coefficient corresponding to each feature in the contextual features in turn. A weight threshold is preset, and each weight coefficient is compared with the weight threshold. Features with a weight coefficient greater than or equal to the weight threshold are considered key contextual factors, while features with a weight coefficient less than the weight threshold are not considered key contextual factors. The method for forming a user context profile includes: Replace the categorical data in the key contextual factors with the corresponding data labels, and combine each key contextual factor and the corresponding weight coefficient into a set of factor vectors, with the factor vectors corresponding to the key contextual factors one-to-one; based on predefined rules, use the rule mapping method to semantically encode each key contextual factor and obtain the semantic label corresponding to each key contextual factor; form a user contextual profile based on the factor vectors, semantic labels, and real-time status labels.

6. The multimodal data-driven intelligent advertising delivery method according to claim 5, characterized in that: Methods for assessing the cost of disruption and the probability of response include: A pre-trained word embedding model is used to convert the semantic labels in the user context profile into corresponding semantic vectors. The converted user context profile is input into a trained indicator prediction model to predict the corresponding user response indicator. The indicator prediction model is a deep neural network model. User response indicators include positive response indicators and negative response indicators. Positive response indicators include click-through rate and conversion rate, and negative response indicators include close rate and bounce rate. A fuzzy inference algorithm is used to predict the response probability range based on the positive response indicator, and the interruption cost range based on the negative response indicator. An optimization algorithm is used to search for the first optimal value from the response probability range as the response probability, and the second optimal value from the interruption cost range as the interruption cost. Methods for determining whether the current time is suitable for delivery include: A preset ratio set includes a proportional coefficient corresponding to the interruption cost and the response probability; the response probability and the interruption cost are multiplied by the corresponding proportional coefficient in the ratio set respectively to obtain the response ratio and the interruption ratio, and the difference between the response ratio and the interruption ratio is used as the response benefit ratio; obtain Historical earnings ratio, The historical return ratio is the most recent one to the current time The response-return ratio corresponding to the historical appropriate delivery time, is an integer greater than 1; the preset factor set includes weighting factors, The weight factors increase in sequence; obtain the historical time corresponding to each historical return ratio, and The historical return ratio and the factor set weight factors to correspond; Each historical return ratio is multiplied by the corresponding weight factor in the factor set, and the average is processed to obtain the return ratio threshold; the response return ratio is compared with the return ratio threshold. If the response return ratio is greater than or equal to the return ratio threshold, the current time is regarded as the appropriate time for delivery. If the response return ratio is less than the return ratio threshold, the current time is not regarded as the appropriate time for delivery.

7. The multimodal data driven intelligent advertising delivery method according to claim 6, characterized in that: The step of searching for a first optimal value from the response probability range as the response probability includes: Step S101: Preset the division coefficient , ; Step S102: Based on the division coefficient , divide the response probability range into two split points and ; Step S103: Calculate two segmentation points respectively and The corresponding matching coefficient will split the points The corresponding matching coefficient is marked as the first coefficient, and the split point The corresponding matching coefficient is marked as the second coefficient; Step S104: Compare the first coefficient with the second coefficient; if the first coefficient is greater than or equal to the second coefficient, proceed to step S105; if the first coefficient is less than the second coefficient, proceed to step S106; Step S105: Calculate the matching coefficient corresponding to the maximum value of the response probability range and mark it as the third coefficient; update the maximum value of the response probability range as the split point , the minimum value of the response probability range remains unchanged; the third coefficient is subtracted from the second coefficient to obtain the coefficient difference, and then proceed to step S107; Step S106: Calculate the matching coefficient corresponding to the minimum value of the response probability range and mark it as the fourth coefficient, and update the minimum value of the response probability range as the split point , the maximum value of the response probability range remains unchanged; subtract the fourth coefficient from the first coefficient to obtain the coefficient difference, and proceed to step S107; Step S107: Determine whether the segmentation is completed. If the segmentation is completed, proceed to step S108. If the segmentation is not completed, update the segmentation coefficient according to the coefficient difference, and re-divide two segmentation points from the response probability range according to the updated segmentation coefficient. and , and returns to step S103; Step S108: Obtain a first optimal value based on the average of the maximum value and the minimum value corresponding to the response probability range, and use it as the response probability; The step of searching for the second optimal value from the interruption cost range as the interruption cost is consistent with the step of searching for the first optimal value from the response probability range as the response probability.

8. The multimodal data-driven intelligent advertising delivery method according to claim 7, characterized in that: In step S102, the segmentation point The expression is: ; Split point The expression is: ;in, is the minimum value of the response probability range, is the maximum value of the response probability range; In step S107, the method for determining whether the segmentation is completed includes: A width threshold is preset, and the width threshold is a real number greater than 0; the maximum value of the updated response probability range is subtracted from the corresponding minimum value to obtain the interval width; the interval width is compared with the width threshold; if the interval width is greater than or equal to the width threshold, the segmentation is not completed; if the interval width is less than the width threshold, the segmentation is completed; Methods for updating the split coefficients based on coefficient differences include: Preset adjustment coefficient and adjustment threshold, the adjustment coefficient is The adjustment threshold is an integer greater than 0; the coefficient difference is compared with the adjustment threshold; if the coefficient difference is greater than or equal to the adjustment threshold, the coefficient difference is multiplied by the adjustment coefficient, and the division coefficient is added to obtain a first adjustment coefficient, and the division coefficient is updated according to the first adjustment coefficient; if the coefficient difference is less than the adjustment threshold, the coefficient difference is multiplied by the adjustment coefficient to obtain an adjustment component, the division coefficient is subtracted from the adjustment component to obtain a second adjustment coefficient, and the division coefficient is updated according to the second adjustment coefficient.

9. The multimodal data-driven intelligent advertising delivery method according to claim 8, characterized in that: The response behavior data includes click behavior, conversion behavior, closing behavior and bounce behavior; When the response behavior data is click behavior or conversion behavior, the timing evaluation system will not be dynamically optimized; When the response behavior data is a close behavior or a bounce behavior, the methods for dynamically optimizing the timing evaluation system include: Preset the optimization factor, add the product of the split coefficient and the optimization factor to the split coefficient to obtain the optimization coefficient; Subtract the product of the width threshold and the optimization factor from the width threshold to obtain the optimization threshold; obtain the coefficient optimization range according to the segmentation coefficient and the optimization coefficient, that is, the segmentation coefficient is the minimum value of the coefficient optimization range, and the optimization coefficient is the maximum value of the coefficient optimization range; obtain the threshold optimization range according to the width threshold and the optimization threshold, that is, the width threshold is the maximum value of the threshold optimization range, and the optimization threshold is the minimum value of the threshold optimization range; randomly select a value from each of the coefficient optimization range and the threshold optimization range, and use them as a set of optimization sets to obtain a total of Group optimization collection, is an integer greater than 1; According to each group of optimization sets, the response probability and interruption cost corresponding to each group of optimization sets are recalculated; the matching coefficient corresponding to the response probability of each group of optimization sets is calculated and marked as the forward coefficient; the matching coefficient corresponding to the interruption cost of each group of optimization sets is calculated and marked as the reverse coefficient; the forward coefficient and reverse coefficient of each group of optimization sets are averaged to obtain the coefficient mean; all coefficient means are compared separately, and the optimization set with the largest coefficient mean is marked as the optimal set; the segmentation coefficient and width threshold are dynamically optimized according to the values ​​in the optimal set.

10. A multimodal data driven intelligent advertising delivery system, implementing a multimodal data driven intelligent advertising delivery method according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, used to obtain users' historical behavior data; The activity analysis module integrates time series pattern mining technology and clustering algorithms to perform time series analysis on historical behavior data, extract the periodic patterns of user activity, and identify candidate delivery windows using the Markov decision process. The data collection module is used to detect the matching status of the current time and the candidate delivery window in real time. When it is detected that the current time matches the candidate delivery window, the real-time behavior data is collected; The context building module is used to extract features from real-time behavioral data, obtain context features, and use the attention mechanism to identify key context factors within the context features to form a user context profile; The timing decision module is used to dynamically evaluate the interruption cost and response probability based on the user's contextual profile through a pre-built timing evaluation system, and to determine whether the current time is suitable for delivery. If so, the ad is delivered; The feedback optimization module is used to collect the user's response behavior data after receiving the advertisement, and dynamically optimize the timing evaluation system based on the response behavior data.

Images