A method and system for generating dynamic advertising delivery strategies based on user behavior

Abstract

The present invention provides a method and system for generating a dynamic advertising delivery strategy based on user behavior, and relates to the technical field of advertising delivery. The method includes the steps of: constructing a user behavior sequence; determining the correlation between the behavior pattern and the sensitive attribute group; based on the correlation and the current advertising delivery strategy, evaluating the probability of different sensitive attribute groups receiving a specific advertisement, and obtaining a quantitative ambiguity risk measurement by calculating the difference index of the advertisement reception probability between groups; using the ambiguity risk measurement as a constraint condition, adjusting the rules for advertising delivery, and obtaining an adjusted advertising delivery strategy; and delivering advertisements according to the adjusted advertising delivery strategy. The method of the present invention solves the technical problem of how to avoid ambiguity when generating an advertising delivery strategy using a dynamic user behavior sequence and associated contextual information, significantly reducing the discrimination exclusion of users of a specific age group, and improving the fairness of delivery.

Description

Technical Field

[0001] The present invention relates to the technical field of advertisement delivery, and in particular to a method and system for generating a dynamic advertisement delivery strategy based on user behavior. Background Art

[0002] In the modern digital advertising landscape, one of the core functions of a platform is to generate effective advertising strategies based on user behavior data. The platform receives data streams of user behavior across digital touchpoints, such as browsing product detail pages, liking or sharing content, searching for keywords, and engaging within an app. Each behavior record is captured and stored by the system as a time series, containing contextual information such as the behavior type, timestamp, digital environment, and device type.

[0003] The system leverages these dynamic user behavior sequences and their associated contextual information, applying techniques such as sequence analysis, pattern recognition, and machine learning to gain a deep understanding of users' immediate intent, fluctuations in interest, and behavioral patterns. Based on these extracted intents and patterns, the system generates highly targeted advertising strategies, determining which types of ads should be shown to users the next time they appear in a particular ad slot, or which user groups should be prioritized. The core goal of these strategies is to maximize click-through and conversion rates by improving the match between ads and users.

[0004] However, this process of relying heavily on dynamic user behavior sequences and contextual information to generate refined delivery strategies carries the risk of potentially discriminatory delivery. Even if the system is designed to avoid directly using user sensitive attributes, certain specific behavior sequence patterns or contextual information may be statistically significantly associated with groups with these sensitive attributes. When the system targets ads based on these behavior sequence patterns and contextual information indirectly associated with sensitive attributes, even if it improves ad effectiveness in the short term, it may result in other user groups related to sensitive attributes that are less strongly associated with these behavior patterns or contexts but are equally likely to be interested in ads being systematically excluded or having their chances of receiving relevant ads significantly reduced. This difference in ad reception opportunities or quality for different sensitive attribute groups, resulting from statistical associations based on dynamic behavior sequences and contexts, constitutes a potential form of discriminatory delivery.

[0005] Even more challenging is that this heterogeneity isn't a static state; it constantly evolves as user behavior continues to evolve, their contextual environment changes, and the system's delivery strategies adjust dynamically. A user's behavior today might cause the system to identify them as a certain group and receive a specific type of ad, while changes in their behavior tomorrow could alter their profile and the type of ad they receive. At the same time, the system's delivery strategies themselves influence users' subsequent behavior, creating a complex feedback loop. This feedback loop can inadvertently reinforce or alter the statistical distribution of certain behavioral patterns across different user groups, thereby exacerbating or changing the strength of the association between behavioral patterns and sensitive attributes, ultimately exacerbating or changing the form and extent of heterogeneous associations.

[0006] Identifying and quantifying this time-evolving heterogeneity arising from dynamic user behavior sequences and contextual associations presents a significant technical challenge. Traditional heterogeneity assessment methods based on static user profiles struggle to capture this dynamic nature. During the generation or adjustment of ad delivery strategies, it is necessary to assess, in real time or near real time, the potential impact of current delivery decisions based on behavior sequences and context on the likelihood or quality of ad reception for different user groups. This requires the system to continuously analyze the dynamic statistical correlation strength between behavior sequence patterns and contextual information and sensitive attribute groups, predict the potential differences in delivery between different groups caused by these correlations under different delivery strategies, and quantify the resulting heterogeneity risk. Furthermore, after identifying potential dynamic heterogeneity, the system must adjust the strategy without significantly compromising overall ad delivery effectiveness. Such adjustments may involve modifying the weights or logic of targeting rules based on behavior sequences or introducing fairness constraints when dynamically allocating ad placements. This adjustment process requires complex techniques to balance fairness and ad effectiveness objectives to find the optimal balance.

[0007] Therefore, a new method is urgently needed to solve the technical problems of ambiguity identification, quantification and strategy adjustment based on dynamic behavior sequences and contextual information. Summary of the Invention

[0008] The purpose of the present invention is to provide a method and system for generating dynamic advertising delivery strategies based on user behavior, which solves the technical problem of how to avoid discriminatory behavior when generating advertising delivery strategies using dynamic user behavior sequences and associated contextual information, and automatically adjusts the strategy when the risk exceeds the standard, thereby significantly reducing the discriminatory exclusion of users of specific age groups while ensuring the effectiveness of advertising, and improving the fairness of delivery.

[0009] In a first aspect, the present invention provides a method for generating a dynamic advertising delivery strategy based on user behavior, comprising the following steps:

[0010] By acquiring user behavior data and associated context information, a user behavior sequence is constructed; the user behavior sequence is used to record user behavior trajectories;

[0011] Analyze user behavior sequences and contextual information, identify behavior patterns and contextual features that are statistically associated with sensitive attribute groups, and determine the association between behavior patterns and sensitive attribute groups. The association relationship is used to dynamically reflect the strength of the association between behavior patterns and sensitive attribute groups.

[0012] Based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy, the probability of different sensitive attribute groups receiving specific ads is evaluated. By calculating the difference index of ad reception probability between groups, a quantitative discrimination risk metric is obtained. The discrimination risk metric is used to quantify the impact that advertising delivery strategies may cause on the unequal ad reception opportunities of different user groups.

[0013] Using the discriminatory risk metric as a constraint, the advertising delivery rules are adjusted to meet the preset discriminatory risk limit, resulting in an adjusted advertising delivery strategy.

[0014] Carry out advertising delivery according to the adjusted advertising delivery strategy.

[0015] The method for generating a dynamic advertising delivery strategy based on user behavior provided by the present invention, in a digital advertising delivery scenario that processes continuously generated and dynamically changing user behavior sequences and associated contextual information flows, identifies and quantifies in real time or near real time the potential risk of discrimination (i.e., the difference in advertising reception opportunities or quality for different user groups) caused by the statistical correlation between these dynamic behavior patterns and contextual information and sensitive attribute groups that evolves over time. On this basis, the advertising delivery strategy is dynamically adjusted to mitigate this risk of discrimination without significantly sacrificing the overall advertising delivery effect.

[0016] In a second aspect, the present invention provides a system for generating a dynamic advertising delivery strategy based on user behavior, comprising:

[0017] The construction module is used to construct a user behavior sequence by obtaining user behavior data and associated context information; the user behavior sequence is used to record the user behavior trajectory;

[0018] The recognition module is used to analyze user behavior sequences and contextual information. By identifying behavioral patterns and contextual features that are statistically associated with sensitive attribute groups, it determines the correlation between the behavioral patterns and sensitive attribute groups. The correlation relationship is used to dynamically reflect the correlation strength between the behavioral patterns and sensitive attribute groups.

[0019] The evaluation module is used to assess the probability of different sensitive attribute groups receiving specific ads based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy. It also calculates the difference in ad reception probabilities between groups to derive a quantitative disparity risk metric. This disparity risk metric is used to quantify the impact that advertising delivery strategies may have on the unequal ad reception opportunities for different user groups.

[0020] An adjustment module is used to use the difference risk measurement as a constraint condition to adjust the advertising rules to meet the preset difference risk limit and obtain an adjusted advertising strategy;

[0021] The control module is used to deliver advertisements according to the adjusted advertisement delivery strategy.

[0022] As can be seen from the above, the method for generating a dynamic advertising delivery strategy based on user behavior provided by the present invention constructs an advertising delivery strategy generation method based on dynamic statistical association analysis and constraint optimization. The method receives a user behavior data stream and continuously analyzes the statistical association between dynamic user behavior sequence patterns and associated context information and sensitive attribute groups. Based on the dynamic association relationship and the current strategy, the differences in advertising reception opportunities or quality for different sensitive attribute groups are quantified in real time or near real time, and used as a discrimination risk indicator. When generating or adjusting the advertising delivery strategy, the system uses the quantified discrimination risk as a constraint or optimization target, and optimizes the advertising delivery effect by adjusting the strategy parameters while meeting the discrimination risk limit.

[0023] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the embodiments of the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flowchart of a method for generating a dynamic advertisement delivery strategy based on user behavior provided by an embodiment of the present invention.

[0025] Figure 2 A schematic diagram of the structure of a system for generating a dynamic advertisement delivery strategy based on user behavior provided by an embodiment of the present invention.

[0026] Description of labels:

[0027] 100, construction module; 200, identification module; 300, evaluation module; 400, adjustment module; 500, control module; 600, privacy module. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0029] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.

[0030] Reference Attachment Figure 1 The present invention provides a method for generating a dynamic advertising delivery strategy based on user behavior, comprising the following steps:

[0031] By acquiring user behavior data and associated context information, a user behavior sequence is constructed; the user behavior sequence is used to record user behavior trajectories;

[0032] Analyze user behavior sequences and contextual information, identify behavior patterns and contextual features that are statistically correlated with sensitive attribute groups, and determine the association between behavior patterns and sensitive attribute groups. The association is used to dynamically reflect the strength of the association between behavior patterns and sensitive attribute groups.

[0033] Based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy, the probability of different sensitive attribute groups receiving specific ads is evaluated. By calculating the difference index of ad reception probability between groups, a quantitative discrimination risk metric is obtained. The discrimination risk metric is used to quantify the impact that advertising delivery strategies may cause on the unequal ad reception opportunities of different user groups.

[0034] Using the discrimination risk metric as a constraint, the advertising rules are adjusted to meet the preset discrimination risk limit. The adjusted advertising strategy is used to alleviate discrimination against specific user groups to a certain extent while ensuring the overall advertising effect.

[0035] Advertisements are delivered according to the adjusted ad delivery strategy to reduce discriminatory delivery to specific user groups and improve the fairness of ad delivery.

[0036] The method solves the problem of unequal advertising receiving opportunities for different user groups that may be caused by dynamic advertising delivery strategies based on user behavior through a series of steps.

[0037] First, user behavior data and context information are obtained to construct a user behavior sequence. The user behavior sequence records the user's behavior trajectory and provides basic data for analysis.

[0038] Next, we analyze user behavior sequences and contextual information to identify behavioral patterns and contextual features that are statistically correlated with sensitive attribute groups, and determine the association between these patterns. This association reflects the strength of the association between the behavioral patterns and the sensitive attribute groups, revealing potential sources of divergent associations.

[0039] Then, based on the identified associations and the current advertising strategy, the probability of receiving a specific ad for different sensitive attribute groups is assessed, and the difference in ad reception probability between groups is calculated to obtain a quantitative disparity risk metric. The disparity risk metric quantifies the impact of advertising strategies on the unequal ad reception opportunities for different user groups, providing an indicator to measure the degree of disparity.

[0040] Subsequently, using the discriminatory risk metric as a constraint, the ad placement rules are adjusted to meet the pre-set discriminatory risk limit, resulting in an adjusted ad placement strategy. This step utilizes quantified discriminatory risk information to guide strategy adjustments, mitigating discriminatory placement for specific user groups while maintaining overall ad placement effectiveness.

[0041] Finally, ads are delivered according to the adjusted ad delivery strategy. This step implements the adjusted strategy to reduce discriminatory delivery to specific user groups and improve fairness in ad delivery.

[0042] The entire process, from data acquisition, potential association identification, discrepancy quantification to strategy adjustment and implementation, forms a closed loop, systematically solving the discrepancy problem in dynamic advertising based on user behavior.

[0043] Specifically, the system continuously receives behavioral data streams generated by users at digital touchpoints, such as browsing, liking, searching, and dwelling. It then combines this data with contextual information such as timestamps, apps, websites, and devices to construct user behavior sequences. This creates a dynamic record of user behavior trajectories. Furthermore, the system analyzes these behavioral sequences and contextual information, applying statistical methods to identify behavioral patterns or contextual features that are statistically significantly associated with specific sensitive attribute groups (e.g., those based on health status, income level, etc.). For example, it may identify a specific search keyword sequence or browsing behavior within a specific time period as highly correlated with a specific sensitive attribute group. The system then determines the correlation between the behavioral pattern and the sensitive attribute group, which dynamically reflects the strength of the correlation. Based on these correlations and current ad placement rules, the system assesses the probability of different sensitive attribute groups receiving specific ads (e.g., health insurance ads, high-end financial services ads). By calculating the differences in the probability of receiving specific ads between different sensitive attribute groups, the system derives a quantitative disparity risk metric. This metric reflects the degree of disparity in ad exposure across different groups that may result from the current placement strategy. The system then uses this disparity risk metric as an optimization or constraint to adjust ad placement rules. The adjustment process aims to reduce the discriminatory risk metric to meet a pre-set risk threshold while maintaining or optimizing overall ad delivery performance (e.g., click-through rate, conversion rate). The system then generates an adjusted ad delivery strategy. Ultimately, the system delivers ads based on the adjusted strategy. This process reduces discriminatory delivery to groups with sensitive attributes associated with specific behavioral patterns, thereby improving the fairness of ad delivery.

[0044] This technical solution works by constructing a closed-loop dynamic advertising strategy adjustment method. First, it continuously receives and processes real-time user behavior data streams and contextual information to construct dynamic user behavior sequences. Next, it continuously analyzes these behavior sequences and contexts to identify patterns and features statistically associated with sensitive attribute groups (such as age, occupation, and geographic location), and dynamically maintains these associations. Based on these dynamic associations and the currently executing advertising strategy, the disparity risk quantification module predicts, in real time or near real time, the probability of receiving a specific ad or ad category, or the expected quality (such as click-through rate or conversion rate) for different sensitive attribute groups (inferred from their behavior patterns and context). It also calculates the statistical differences in ad reception chance or quality between these groups, using this as a quantitative disparity risk indicator. Subsequently, the quantified disparity risk metric is received. When generating new advertising strategies or fine-tuning existing ones, the disparity risk value is used as a constraint in the optimization process (for example, requiring that differences between different groups do not exceed a certain threshold) or as part of a multi-objective optimization function (for example, minimizing disparity risk while maximizing overall click-through rate). By adjusting strategy parameters (such as the weights of different behavioral patterns, the priority of targeting rules, and bidding factors), the optimal or suboptimal strategy is determined while meeting fairness requirements. Finally, actual ad placement decisions are made based on the optimized strategy. Because user behavior, context, and system policies are constantly changing, this process is dynamic and cyclical. This method can monitor discriminatory risks in real time and make corresponding strategy adjustments, thereby mitigating potential discriminatory placement in a dynamic environment.

[0045] In some specific embodiments, constructing user behavior sequences can involve collecting user behavior logs from multiple sources, such as e-commerce platforms, social media, and search engines, and organizing these logs in chronological order. Each record contains information such as behavior type, time, location, device, and content category. When analyzing behavior sequences and contextual information to determine associations, sequential pattern mining techniques can be used to identify frequent behavior sequences and calculate the frequency or statistical correlation of these sequences across different sensitive attribute groups (for example, using mutual information or a chi-square test). When assessing the probability of different groups receiving a specific ad, the predicted scores for users belonging to different sensitive attribute groups can be calculated based on the current ad delivery model (for example, a model that predicts user clicks or conversions) and the identified association patterns. These scores are then aggregated to obtain group-level probabilities. Inter-group dissimilarity metrics can be calculated using the absolute or relative difference between the average probabilities of different groups. When adjusting ad delivery rules, a discriminatory risk metric can be used as a constraint to adjust or filter the weights of targeting rules based on behavioral patterns to reduce the impact of high-risk association patterns until the discriminatory risk metric falls below a preset threshold of 0.05. For example, if a specific behavior pattern is found to be highly correlated with a certain sensitive attribute group, and the delivery based on this pattern results in the probability of this group receiving a certain type of advertisement being significantly lower than that of other groups, then when adjusting the strategy, the weight of this behavior pattern in determining the delivery of this type of advertisement can be reduced, or additional delivery opportunities can be introduced for this sensitive attribute group to balance the reception probability among different groups.

[0046] In some embodiments, the steps of analyzing user behavior sequences and context information, identifying behavior patterns and context features that are statistically correlated with sensitive attribute groups, and determining the correlation between the behavior patterns and the sensitive attribute groups include:

[0047] For a small sample of sensitive attribute groups, perform the following steps A1-A5:

[0048] A1. Perform standardized preprocessing on user behavior sequences, including missing value filling, noise data filtering, and data format conversion, to obtain standardized behavior sequence data.

[0049] A2. An adaptive sliding window approach is used to extract frequent behavior patterns from standardized behavior sequence data and, combined with contextual features, to construct a set of candidate association rules. The size of the sliding window is adjusted based on the average time interval between user behaviors.

[0050] A3. Using mutual information and chi-square tests, calculate the strength of association between each association rule in the candidate association rule set and the small sample sensitive attribute group. Based on the strength of association between each association rule in the candidate association rule set and the small sample sensitive attribute group, select significant association rules from the candidate association rule set that exceed a preset threshold as identified association rules.

[0051] A4. Use data augmentation technology to generate simulated behavior sequences based on the identified association rules, and use the generated simulated behavior sequences as supplementary user behavior sequences.

[0052] A5. Analyze the supplemented user behavior sequence and contextual information, and determine the correlation between the behavior pattern and the small sample sensitive attribute group by identifying the behavior patterns and contextual features that are statistically associated with the small sample sensitive attribute group.

[0053] In digital advertising, the use of dynamic user behavior sequences and associated contextual information for ad targeting presents challenges such as uneven user behavior data quality, time-varying user behavior patterns, and imbalanced data volumes across sensitive attribute groups, all of which hinder the accuracy and robustness of association identification. This solution addresses these issues through a series of steps. First, the raw user behavior sequences undergo standardized preprocessing. This process improves data quality and consistency through missing value filling, noise filtering, and data format conversion, laying the foundation for subsequent analysis. Improving data quality is particularly important for small sample data, as it reduces the impact of noise on analysis results. Next, an adaptive sliding window approach is used to extract frequent behavior patterns from the preprocessed behavior sequences and, combined with contextual features, construct candidate association rules. The adaptive sliding window size is adjusted based on the average time interval between user behaviors, enabling more flexible capture of behavior patterns across different time spans. This step identifies potential behavior patterns and contextual features, forming a candidate set for analysis. Statistical methods such as mutual information and the chi-square test are then used to calculate the strength of association between the candidate association rules and the small sample sensitive attribute group, screening for significant association rules whose strength exceeds a preset threshold. This step attempts to identify preliminary, statistically significant association rules from a limited, small sample size. However, due to the small sample size, the number of significant rules directly obtained may be limited or unstable. To address the small sample size issue, data augmentation techniques are used to generate simulated behavior sequences based on the identified association rules. These simulated sequences mimic the known behavioral characteristics of the small sample population, effectively increasing the amount of data available for analysis. Finally, association analysis is performed again using the user behavior sequences and contextual information supplemented with the simulated behavior sequences. This increased data volume enhances the robustness of the statistical analysis, enabling more accurate and comprehensive identification of behavioral patterns and contextual features statistically associated with the small sample sensitive attribute population, thereby determining more reliable associations between behavioral patterns and the small sample sensitive attribute population. Through the above steps, this solution, when processing small sample sensitive attribute populations, first performs data preprocessing and preliminary pattern recognition, then utilizes data augmentation techniques to expand the data volume. Finally, a more reliable association analysis is performed on the expanded dataset. This addresses issues such as uneven user behavior data quality, temporal variations in user behavior patterns, and imbalanced data volumes across sensitive attribute populations, thereby improving the accuracy and robustness of association identification.

[0054] In some specific embodiments, for a small sample of users with sensitive attributes representing a specific health condition, their original behavior sequences may contain missing event timestamps or unusual browsing durations. First, standardization preprocessing is performed, such as filling missing timestamps with the previous valid value, smoothing browsing durations using a median filter, and converting all event types into a unified encoding format. This results in standardized behavior sequence data. Next, an adaptive sliding window approach is used to extract frequent behavior patterns from the standardized data based on the average time interval between user behaviors in the group (e.g., if the average interval is 1 hour, the window size is set to 2 hours). For example, frequent behavior patterns (e.g., "searching for a specific disease name -> browsing medical forums -> viewing drug information") are extracted. Combined with contextual features (e.g., access time during weekdays), a set of candidate association rules is constructed. Mutual information and a chi-square test are then used to calculate the strength of association between these candidate rules and the health condition group. Rules with association strengths exceeding a preset threshold are selected, such as those found to be significantly associated with "searching for a specific disease name" in the group. Furthermore, data augmentation techniques are used to generate simulated behavior sequences based on the identified association rules (e.g., "searching for a specific disease name"). Specifically, a behavioral feature space for this group is constructed, the frequency of "searching for a specific disease name" within this group is calculated, and a generative adversarial network (GAN) is used to generate simulated sequences containing this "searching for a specific disease name" behavior. These simulated sequences are then added to the user behavior sequences for this group. Finally, the user behavior sequences and contextual information after the simulated sequences are analyzed to further identify behavioral patterns and contextual features that are statistically associated with this group, thereby determining a more accurate association between the behavioral patterns and this health status group. Through data augmentation, even with limited original small sample data, behavioral patterns associated with this sensitive attribute group can be more reliably identified, improving the accuracy of association identification.

[0055] In some embodiments, the specific steps in step A4 include:

[0056] Determine the behavioral feature space of a small sample sensitive attribute group, which is composed of the behavioral patterns and contextual features involved in the identified association rules;

[0057] For each behavioral feature in the behavioral feature space, calculate the frequency of occurrence of the behavioral feature in the small sample sensitive attribute group to obtain the behavioral feature frequency distribution;

[0058] Based on the frequency distribution of behavioral features, a generative adversarial network (GAN) is used to generate simulated behavior sequences that are similar to the behavior patterns of a small sample of sensitive attribute groups. The input to the generator in the GAN is random noise, and the discriminator is used to distinguish between real behavior sequences and simulated behavior sequences. Through adversarial training, the generator's ability to generate high-quality simulated behavior sequences is improved.

[0059] The generated simulated behavior sequence is added to the user behavior sequence of the small sample sensitive attribute group to form a supplemented user behavior sequence for subsequent association relationship analysis.

[0060] This technical solution aims to solve the problem of how to generate effective and realistic simulated behavior sequences when performing data augmentation on small-sample sensitive attribute groups. First, the behavioral feature space of the small-sample sensitive attribute group is determined. This space is composed of the behavioral patterns and contextual features involved in the identified association rules, thereby focusing on the key behavioral features related to the sensitive attribute group. Next, the frequency of occurrence of each behavioral feature in this behavioral feature space in the existing small-sample sensitive attribute group data is calculated to obtain the behavioral feature frequency distribution, which provides a statistical overview of the behavioral patterns of the small-sample group. A generative adversarial network (GAN) is then used to generate simulated behavior sequences. The generator attempts to generate simulated sequences based on the behavioral feature frequency distribution, and the discriminator is used to distinguish between simulated sequences and real sequences. Through adversarial training, the generator learns to generate simulated sequences with similar statistical features and patterns to real behavior sequences. This overcomes the problem of difficulty in capturing the diversity of behavioral patterns caused by insufficient small-sample data and improves the authenticity of the generated data. Finally, the generated simulated behavior sequence is added to the user behavior sequence of the small sample sensitive attribute group to form a supplemented user behavior sequence. By increasing the amount of data, more sufficient data support is provided for subsequent association analysis, which helps to more accurately identify the association between behavior patterns and small sample sensitive attribute groups, thereby improving the accuracy of association analysis for small sample sensitive attribute groups.

[0061] In some specific embodiments, for a small sample of sensitive attribute users interested in a specific health condition, the amount of user behavior data for this group is relatively small. First, the system uses preliminary analysis to identify behavioral patterns that are statistically correlated with this group, such as searching for specific disease names during a specific time period, browsing specific medical forums, and clicking on specific health information links, as well as related contextual features, such as the categories of websites visited and the types of devices used. These behavioral patterns and contextual features constitute the behavioral feature space for this group. Next, the system counts the frequency of occurrence of these identified behavioral features within this small sample, for example, the frequency of "searching for disease A" is X, and the frequency of "browsing forum B" is Y. Then, using these frequency distributions as guidance, a generative adversarial network is trained. The generator takes random noise as input and attempts to generate simulated behavioral sequences, such as a sequence consisting of "searching for disease A" and "browsing forum B." The discriminator receives the real group behavior sequences and the simulated sequences generated by the generator and attempts to distinguish them. Through adversarial training, the generator learns to generate simulated sequences that resemble the real sequences in terms of statistical features and patterns. For example, the generated sequences tend to include behavioral features that are more frequent in real data and may mimic the order or combination of behavioral features in real sequences. Finally, these generated simulated behavioral sequences are added to the original behavioral sequence dataset for the small sample sensitive attribute group to form an expanded dataset. This expanded dataset is used to subsequently more accurately analyze the strength of the association between behavioral patterns and the sensitive attribute group, such as calculating the mutual information or chi-square value between "search for disease A" and the group, thereby improving the ability to identify key association patterns in data-scarce situations.

[0062] In some embodiments, after determining the association relationship, the method further includes the following steps:

[0063] Differential privacy technology is used to perturb the association rules corresponding to the association relationship to prevent the leakage of sensitive attribute information.

[0064] This technical solution solves the problem that after determining the association relationship between behavior patterns and sensitive attribute groups, association rules may leak sensitive attribute information. After analyzing user behavior sequences and contextual information, identifying the statistical association between behavior patterns and sensitive attribute groups, and determining the association relationship, these association relationships are expressed in the form of association rules. Since these rules reveal the statistical connection between behavior patterns and sensitive attribute groups, there is a privacy risk in directly using or storing them. In order to prevent the leakage of sensitive attribute information, differential privacy technology is used to perturb these association rules. Specifically, by adding calculated random noise to the statistical parameters or form of the association rules, even if an attacker obtains the perturbed association rules, it is difficult to accurately infer the original, undisturbed association information, thereby protecting the privacy of users related to sensitive attributes. As a result, the risk of privacy leakage can be reduced while using this association information to evaluate the risk of divergence and adjust the delivery strategy.

[0065] In some specific embodiments, assume that an association rule is identified: there is an association between the behavior pattern "frequent browsing of health-related products" and the sensitive attribute group "users with specific health conditions," with a confidence level of 0.75. To protect privacy, this confidence level is perturbed using a Laplace mechanism. First, the sensitivity for the confidence calculation is determined, for example, 1 / N, where N is the total number of users. Then, based on a preset privacy budget ε, the scale parameter b of the Laplace distribution is calculated as sensitivity / ε. A random noise value is drawn from a Laplace distribution with mean 0 and scale parameter b. This noise value is added to the original confidence level of 0.75 to obtain the perturbed confidence level of 0.75 + noise. For example, if the noise value is +0.02, the perturbed confidence level is 0.77; if the noise value is -0.03, the perturbed confidence level is 0.72. This perturbed confidence level is used instead of the original 0.75 in subsequent assessments of risk of discrepancy. Therefore, even if an attacker analyzes the confidence after perturbation, it is difficult to accurately infer the original association strength, thereby protecting the privacy information related to the "users with specific health conditions" group.

[0066] In some embodiments, the steps of using differential privacy technology to perturb the association rules corresponding to the association relationship to prevent leakage of sensitive attribute information include:

[0067] B1. Determine the set of association rules to be perturbed and assign an initial privacy budget to each association rule; the sum of the initial privacy budgets must satisfy the preset total privacy budget constraint.

[0068] B2. Evaluate the sensitivity of each association rule by considering the sensitive attribute types involved, the support of the association rule, and the confidence of the association rule;

[0069] B3. Based on the sensitivity of the association rules, adjust the privacy budget allocation for each association rule; association rules with higher sensitivity are allocated smaller privacy budgets, while association rules with lower sensitivity are allocated larger privacy budgets;

[0070] B4. Based on the adjusted privacy budget and the data type of the association rule, select a differential privacy mechanism for each association rule. Differential privacy mechanisms include the Laplace mechanism and the Gaussian mechanism.

[0071] B5. Use the selected differential privacy mechanism to perturb the association rules and generate perturbed association rules;

[0072] B6. Verify the perturbed association rules to ensure that they meet the preset support constraints, confidence constraints, and privacy protection constraints. If the constraints are not met, return to step B3 and readjust the privacy budget allocation.

[0073] The set of association rules to be perturbed is determined, and an initial privacy budget is assigned to each association rule in the set. The sum of the initial privacy budgets is constrained to fall within the preset total privacy budget. This sets the overall privacy protection level. Furthermore, the sensitivity of each association rule is evaluated by considering the sensitive attribute types involved, the support of the association rule, and the confidence of the association rule. Specifically, a sensitive attribute type grading system is constructed, classifying different sensitive attribute types into multiple levels and assigning each level a corresponding grade score. Health-related attributes receive higher grade scores than interest-related attributes. Based on the support of the association rule, the number of users covered by the rule is calculated, and the support score is determined based on this number; the greater the number of covered users, the higher the support score. Based on historical association rule data, the confidence of the association rule is evaluated to obtain a confidence score. The sensitivity score of the association rule is obtained by taking the weighted sum of the grade score, support score, and confidence score. The weights of the support score and confidence score are both lower than those of the grade score to ensure that the sensitive attribute type plays a dominant role in the comprehensive evaluation. Based on the sensitivity assessment results of the association rules, the privacy budget allocation for each association rule is adjusted. High-sensitivity association rules are assigned a smaller privacy budget, while low-sensitivity association rules are assigned a larger privacy budget. This achieves a balance between privacy protection strength and data availability. Based on the adjusted privacy budget and the data type of the association rule, a differential privacy mechanism is selected for each association rule. Differential privacy mechanisms include Laplace and Gaussian mechanisms, and the appropriate mechanism is selected to accommodate different types of data perturbations. Using the selected differential privacy mechanism, the association rules are perturbed to generate perturbed association rules. This introduces noise into the association rules, achieving privacy protection. The perturbed association rules are then verified to ensure that they meet the preset support, confidence, and privacy constraints. If the constraints are not met, the process returns to step B3 and readjusts the privacy budget allocation, forming a feedback loop until all constraints are met.

[0074] Specifically, this technical solution aims to address the problem of balancing privacy protection strength and post-perturbation rule usability when perturbing association rules. First, the set of association rules to be perturbed is identified and an overall privacy budget is set. Then, a sensitivity assessment is performed on each association rule in the set. This sensitivity assessment takes into account the type of sensitive attributes associated with the rule, the user scope covered by the rule, and the reliability of the rule. A sensitive attribute type classification system ensures that the privacy risks associated with different sensitive attribute types are reflected. Support and confidence scores reflect the universality and reliability of the rule. A comprehensive sensitivity score is derived through weighted summation, with sensitive attribute types receiving a higher weight, ensuring that privacy risks dominate. Based on the assessed sensitivity, the privacy budget assigned to each rule is dynamically adjusted. Rules with higher sensitivity are considered to pose a greater privacy risk and are therefore allocated a smaller privacy budget, introducing more noise to provide stronger privacy protection. Rules with lower sensitivity are allocated a larger privacy budget, introducing less noise to preserve more of the rule's original information and maintain usability. This step is key to achieving a balance between privacy protection and data availability. An appropriate differential privacy mechanism is selected based on the rule data type to ensure the effectiveness of the perturbation process. Using the adjusted privacy budget and the selected mechanism, the association rules are actually perturbed to generate noisy rules. Finally, the perturbed rules are verified. This process checks whether the perturbed rules still meet the preset business or technical constraints and privacy protection requirements. If the verification fails, it indicates that the current budget allocation or perturbation may be inappropriate, and the budget needs to be readjusted, forming a feedback loop until a perturbation result that meets all constraints is found. Through the above steps, a structured and adaptive association rule perturbation method is provided. It can adjust the privacy protection strength according to the characteristics of the rules themselves, thereby protecting user privacy while maintaining the practicality of the association rules as much as possible, which can be used for subsequent adjustments to advertising strategies.

[0075] In some specific implementations, assume that two association rules are identified that require perturbation: Rule R1: "Browsing health forums" => "Belongs to the health-sensitive attribute group", with support 0.05 and confidence 0.8; Rule R2: "Browsing food blogs" => "Belongs to the interest-sensitive attribute group", with support 0.1 and confidence 0.7. The preset total privacy budget is Epsilon = 1.0.

[0076] First, the set of association rules to be disturbed is determined to be {R1, R2}. The initial privacy budget is allocated to each rule as 0.5.

[0077] Next, evaluate the sensitivity of each rule. Assume that the health attribute is rated 5, and the interest attribute is rated 2. The support score is calculated based on the number of covered users. Assume that a support score of 0.05 corresponds to a score of 3, and a support score of 0.1 corresponds to a score of 4. The confidence score directly uses the confidence value, with R1 of 0.8 and R2 of 0.7. Assume that the weight of the grade score is 0.6, the weight of the support score is 0.2, and the weight of the confidence score is 0.2.

[0078] Rule R1 sensitivity score = 5*0.6+3*0.2+0.8*0.2=3.0+0.6+0.16=3.76.

[0079] Rule R2 sensitivity score = 2*0.6+4*0.2+0.7*0.2=1.2+0.8+0.14=2.14.

[0080] Adjust the privacy budget based on sensitivity. The total sensitivity is 3.76 + 2.14 = 5.9.

[0081] The adjusted budget for Rule R1 = Epsilon*(1-3.76 / 5.9)=1.0*(1-0.637)≈0.363.

[0082] Rule R2 adjusted budget = Epsilon*(1-2.14 / 5.9)=1.0*(1-0.363)≈0.637.

[0083] R1 with high sensitivity is allocated a smaller budget, and R2 with low sensitivity is allocated a larger budget.

[0084] Assuming that the confidence of the association rule is a numerical type, the Laplace mechanism is selected for perturbation. The noise intensity of the Laplace mechanism is inversely proportional to the privacy budget.

[0085] Use the adjusted budget to perturb the confidence of the rule. For example, add noise to the confidence of R1 at 0.8, with the noise level determined by the budget of 0.363. Add noise to the confidence of R2 at 0.7, with the noise level determined by the budget of 0.637. Generate the perturbed confidence value.

[0086] Verify the association rules after perturbation. For example, check whether the confidence after perturbation is still greater than a preset minimum confidence threshold (e.g., 0.5). If the confidence of R1 after perturbation is less than 0.5, the verification fails and the process returns to step B3. It may be necessary to fine-tune the budget allocation between R1 and R2, for example, slightly increasing the budget for R1 and reducing the budget for R2, and then re-perturb and verify until the constraints are met.

[0087] It should be noted that in practical applications, data such as support and confidence in association rules are all numerical data. For numerical data, the Laplace and Gaussian mechanisms are generally used. The Laplace mechanism achieves differential privacy by adding noise that follows a Laplace distribution to the data, and is suitable for perturbing query results. The Gaussian mechanism, on the other hand, adds noise that follows a Gaussian distribution and performs well with high-dimensional data and complex queries.

[0088] In some embodiments, the specific steps in step B2 include:

[0089] Construct a sensitive attribute type grading system. The sensitive attribute type grading system divides different sensitive attribute types into multiple levels and assigns a corresponding grade score to each level; health-related attributes are scored higher than interest-related attributes;

[0090] According to the support of the association rule, the number of users covered by the rule is calculated, and the support score is determined based on the number of users; the more users covered, the higher the support score;

[0091] Based on historical association rule data, the confidence of the association rule is evaluated to obtain a confidence score;

[0092] The sensitivity score of the association rule is obtained by weighted summing the grade score, support score, and confidence score. The weights of the support score and confidence score are lower than the weight of the grade score to ensure that the sensitive attribute type plays a dominant role in the comprehensive evaluation.

[0093] A sensitive attribute type grading system was constructed, which categorizes different sensitive attribute types into different levels and assigns a grade score to each level. For example, health-related attributes are assigned a higher grade score than interest-related attributes. This quantifies the inherent differences in sensitivity between different sensitive attribute types. Furthermore, based on the support of the association rule, the number of users covered by the rule is calculated, and the support score is determined based on this number. The greater the number of covered users, the higher the support score, reflecting the scope of the rule's impact. Simultaneously, based on historical association rule data, the confidence of the association rule is evaluated to obtain a confidence score, which reflects the predictive strength of the rule. Finally, the grade score, support score, and confidence score are weighted and summed to calculate the sensitivity score of the association rule. During the weighting process, the support score and confidence score are weighted lower than the grade score to ensure that the sensitive attribute type has a greater impact in the comprehensive sensitivity assessment.

[0094] Specifically, this technical solution provides a method for quantitatively assessing the sensitivity of association rules by comprehensively considering the sensitive attribute types involved in the association rule, the number of users covered by the rule, and the confidence level of the rule. First, a sensitive attribute type grading system is established, assigning different grading scores to different sensitive attribute types. For example, health-related attributes are given higher grading scores than interest-related attributes. This distinguishes the inherent sensitivity of different sensitive attribute types. Next, the number of users covered by the association rule is calculated based on its support level and converted into a support score. A higher support score indicates the user range affected by the rule. The confidence level of the association rule is then evaluated based on historical data to obtain a confidence score, which reflects the predictive accuracy of the rule. Finally, these three scores are weighted and summed to obtain a comprehensive sensitivity score for the association rule. During the weighting process, the weight of the sensitive attribute type grading score is set higher than that of the support and confidence scores, ensuring that the sensitive attribute type plays a dominant role in the final sensitivity assessment. Through this comprehensive and weighted approach, a more accurate association rule sensitivity measurement can be obtained to guide the privacy budget allocation of the subsequent differential privacy mechanism, thereby solving the problem of unreasonable privacy budget allocation caused by inaccurate association rule sensitivity assessment, thereby improving privacy protection effects and balancing data availability.

[0095] In certain embodiments, the steps of evaluating the probability of different sensitive attribute groups receiving a specific advertisement based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy, and obtaining a quantitative disparity risk metric by calculating a difference index of advertisement reception probability between groups include:

[0096] Based on the association relationships, multiple behavioral patterns with the highest association strength with each sensitive attribute group are determined, and a set of sensitive attribute behavioral patterns is constructed;

[0097] Based on the current advertising delivery strategy and the set of sensitive attribute behavior patterns, the probability of each user in each sensitive attribute group receiving a specific advertisement is calculated to obtain the first advertisement reception probability representing the user level;

[0098] For each sensitive attribute group, the first advertisement reception probability of all users in the corresponding sensitive attribute group is summarized, and the group average advertisement reception probability is calculated to obtain the second advertisement reception probability representing the group level;

[0099] For every two sensitive attribute groups, calculate the absolute value of the difference between the second advertisement reception probabilities corresponding to the two groups and use it as the corresponding difference index;

[0100] The difference indicators between all sensitive attribute groups are weighted averaged to obtain the divergence risk measurement; the weight of each difference indicator is proportional to the population size of the corresponding two sensitive attribute groups, and the larger the population size, the greater the weight.

[0101] This method leverages the correlation between identified behavioral patterns and sensitive attribute groups to identify sets of behavioral patterns with a high degree of association with each sensitive attribute group. These behavioral patterns reflect the characteristics of that group. Next, based on the current advertising strategy, the system calculates the probability of each user within each sensitive attribute group receiving a specific ad. This probability takes into account whether the user's specific behavior conforms to the targeting rules of the current strategy and the association between that behavior and the sensitive attribute group. These user-level probabilities are aggregated to calculate the average ad reception probability for each sensitive attribute group. This group-level average probability represents the overall chance of receiving a specific ad for that group. To quantify disparity, the method calculates the absolute difference in the average ad reception probability between any two sensitive attribute groups. This difference reflects the degree of inequality in the two groups' chances of receiving a specific ad. Finally, the disparity indicators for all pairs of sensitive attribute groups are weighted and averaged to produce a single disparity risk metric. The weights are set proportionally to the group population size, ensuring that differences between large groups are given greater attention, allowing the final disparity risk metric to reflect the impact on a wider user base.

[0102] Specifically, this technical solution addresses the issue of selecting appropriate disparity indicators to more accurately reflect disparity risk through the following steps. First, based on the correlation data obtained in the previous step, which reflects the strength of association between behavioral patterns and sensitive attribute groups, the system identifies sets of behavioral patterns whose correlation strength with each sensitive attribute group reaches a preset threshold or ranks highly. This constructs a sensitive attribute behavior pattern set, which contains key patterns reflecting the behavioral characteristics of each sensitive attribute group. Next, based on the current advertising delivery strategy, which defines the rules and preferences for advertising delivery, and the constructed sensitive attribute behavior pattern set, the system evaluates each user within each sensitive attribute group. The system assesses the probability of a user receiving a specific ad, for example, by analyzing whether the user's behavior sequence contains patterns in the sensitive attribute behavior pattern set and how the current delivery strategy responds to these patterns. This results in a first ad reception probability representing the user level. Furthermore, for each sensitive attribute group, the first ad reception probabilities for all users within that group are aggregated, for example by calculating the arithmetic mean, to obtain a second ad reception probability representing the group level. This probability reflects the likelihood that the sensitive attribute group as a whole will receive the specific ad. Next, to quantify the differences between different groups, the absolute value of the difference in the probability of receiving the second ad is calculated for any two sensitive attribute groups. This absolute difference serves as the corresponding difference indicator, directly reflecting the degree of inequality in the two groups' chances of receiving a specific ad. Finally, to obtain an overall disparity risk metric, the difference indicators for all possible pairs of sensitive attribute groups are weighted and averaged. The weight of each difference indicator is proportional to the population size of the two sensitive attribute groups, with the larger the population, the greater the weight. Therefore, differences between groups with larger populations contribute more to the final disparity risk metric, allowing the metric to more accurately reflect the impact on a wider range of user groups. Through this series of steps, the solution provides a quantitative process from individual behavior probabilities to group differences and then to an overall risk metric. It addresses the issue of how to select appropriate difference indicators to more accurately reflect disparity risk. Different difference indicators may be sensitive to different disparity situations, and the most appropriate indicator needs to be selected based on the specific business scenario.

[0103] In some specific implementations, assume that an advertising platform focuses on age-sensitive attributes and divides users into three groups: Group 1 (20-30 years old), Group 2 (30-50 years old), and Group 3 (50-60 years old). Using the steps of the above method, the average probability of receiving a specific advertisement for each of these three groups is calculated to be 0.1 for Group 1, 0.3 for Group 2, and 0.15 for Group 3. Furthermore, the populations of these three groups are 1,000, 2,000, and 500, respectively.

[0104] According to the above calculation method:

[0105] 1. Calculate the difference index between groups (absolute value of the difference):

[0106] The difference index between group 1 and group 2: |0.1-0.3|=0.2;

[0107] The difference index between group 1 and group 3: |0.1-0.15|=0.05;

[0108] The difference index between group 2 and group 3: |0.3-0.15|=0.15;

[0109] 2. Calculate the weight of the weighted average: The weight is proportional to the population size of the two groups. One specific implementation method can be to equal the weight to the sum of the population size of the two groups.

[0110] The weight of group 1 and group 2: 1000 + 2000 = 3000;

[0111] The weight of group 1 and group 3: 1000 + 500 = 1500;

[0112] The weight of group 2 and group 3: 2000 + 500 = 2500;

[0113] Total weight: 3000+1500+2500=7000;

[0114] 3. Calculate the quantitative divergence risk metric (weighted average divergence index):

[0115] Dissimilar risk measure = (0.2*3000+0.05*1500+0.15*2500) / 7000 = (600+75+375) / 7000 = 1050 / 7000 = 0.15;

[0116] During this process, this feature calculates the absolute differences in ad reception probabilities between all group pairs and weights these differences by group population size to produce a single value (0.15). This value quantifies the age-related risk of bias caused by the current ad delivery strategy. This calculation method considers the magnitude of differences between groups and assigns higher weights to group pairs with larger populations, thereby more accurately reflecting the actual user range and degree of bias. This quantified bias risk metric (0.15) is then used in policy constraint optimization to guide the system in adjusting delivery rules to reduce this value and mitigate bias risk.

[0117] Reference Attachment Figure 2 The present invention provides a system for generating a dynamic advertising delivery strategy based on user behavior, comprising:

[0118] The construction module 100 is used to construct a user behavior sequence by acquiring user behavior data and associated context information; the user behavior sequence is used to record the user behavior trajectory;

[0119] Identification module 200 is used to analyze user behavior sequences and context information, identify behavior patterns and context features that are statistically correlated with sensitive attribute groups, and determine the correlation between the behavior patterns and sensitive attribute groups; the correlation relationship is used to dynamically reflect the correlation strength between the behavior patterns and sensitive attribute groups;

[0120] Evaluation module 300 is used to evaluate the probability of different sensitive attribute groups receiving a specific advertisement based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy. It also calculates the difference index of the ad reception probability between groups to obtain a quantitative disparity risk metric. The disparity risk metric is used to quantify the impact of the advertising delivery strategy on the unequal ad reception opportunities of different user groups.

[0121] An adjustment module 400 is configured to use the ambiguity risk metric as a constraint to adjust the advertising delivery rules to meet the preset ambiguity risk limit, thereby obtaining an adjusted advertising delivery strategy.

[0122] The control module 500 is configured to deliver advertisements according to the adjusted advertisement delivery strategy.

[0123] In some embodiments, a privacy module 600 is further included, which is used to use differential privacy technology to perturb the association rules corresponding to the association relationship after determining the association relationship, so as to prevent the leakage of sensitive attribute information.

[0124] The foregoing description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for generating a dynamic advertising delivery strategy based on user behavior, characterized in that: The following steps are involved: Construct user behavior sequences by acquiring user behavior data and associated context information; User behavior sequence is used to record user behavior trajectory; Analyze user behavior sequences and contextual information, identify behavior patterns and contextual features that are statistically correlated with sensitive attribute groups, and determine the correlation between behavior patterns and sensitive attribute groups; The association relationship is used to dynamically reflect the association strength between the behavior pattern and the sensitive attribute group; Based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy, the probability of different sensitive attribute groups receiving specific ads is evaluated. By calculating the difference index of ad reception probability between groups, a quantitative discrimination risk metric is obtained. The discrimination risk metric is used to quantify the impact that advertising delivery strategies may cause on the unequal ad reception opportunities of different user groups. Using the discriminatory risk metric as a constraint, the advertising delivery rules are adjusted to meet the preset discriminatory risk limit, resulting in an adjusted advertising delivery strategy. Place advertisements according to the adjusted advertising strategy; Analyzing user behavior sequences and contextual information, identifying behavior patterns and contextual features that are statistically correlated with sensitive attribute groups, and determining the correlation between behavior patterns and sensitive attribute groups includes the following steps: For a small sample of sensitive attribute groups, perform the following steps A1-A5: A1. Perform standardization preprocessing on user behavior sequences to obtain standardized behavior sequence data; A2. Adopting an adaptive sliding window approach, we extract frequent behavior patterns from standardized behavior sequence data and, combined with contextual features, construct a set of candidate association rules. A3. Using mutual information and chi-square tests, calculate the strength of association between each association rule in the candidate association rule set and the small sample sensitive attribute group. Based on the strength of association between each association rule in the candidate association rule set and the small sample sensitive attribute group, select significant association rules from the candidate association rule set that exceed a preset threshold as identified association rules. A4. Use data augmentation technology to generate simulated behavior sequences based on the identified association rules, and use the generated simulated behavior sequences as supplementary user behavior sequences. A5. Analyze the supplemented user behavior sequence and contextual information, identify behavioral patterns and contextual features that are statistically correlated with the small sample sensitive attribute group, and determine the correlation between the behavioral patterns and the small sample sensitive attribute group; The specific steps in step A4 include: Determine the behavioral feature space of a small sample sensitive attribute group, which is composed of the behavioral patterns and contextual features involved in the identified association rules; For each behavioral feature in the behavioral feature space, calculate the frequency of occurrence of the behavioral feature in the small sample sensitive attribute group to obtain the behavioral feature frequency distribution; Based on the frequency distribution of behavioral features, a generative adversarial network is used to generate simulated behavioral sequences similar to the behavioral patterns of a small sample group with sensitive attributes. The generated simulated behavior sequence is added to the user behavior sequence of the small sample sensitive attribute group to form a supplemented user behavior sequence for subsequent correlation analysis; Based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising strategy, the probability of different sensitive attribute groups receiving a specific advertisement is evaluated. The difference index of the ad reception probability between groups is calculated to obtain a quantitative disparity risk measurement. The steps include: Based on the association relationships, multiple behavioral patterns with the highest association strength with each sensitive attribute group are determined, and a set of sensitive attribute behavioral patterns is constructed; Based on the current advertising delivery strategy and the set of sensitive attribute behavior patterns, the probability of each user in each sensitive attribute group receiving a specific advertisement is calculated to obtain the first advertisement reception probability representing the user level; For each sensitive attribute group, the first advertisement reception probability of all users in the corresponding sensitive attribute group is summarized, and the group average advertisement reception probability is calculated to obtain the second advertisement reception probability representing the group level; For every two sensitive attribute groups, calculate the absolute value of the difference between the second advertisement reception probabilities corresponding to the two groups and use it as the corresponding difference index; The difference indicators between all sensitive attribute groups are weighted averaged to obtain the divergence risk measurement; the weight of each difference indicator is proportional to the population size of the corresponding two sensitive attribute groups, and the larger the population size, the greater the weight.

2. The method for generating a dynamic advertising delivery strategy based on user behavior according to claim 1, characterized in that: Standardization preprocessing includes filling missing values, filtering noisy data, and data format conversion.

3. The method for generating a dynamic advertising delivery strategy based on user behavior according to claim 1, characterized in that: When the adaptive sliding window method is adopted, the size of the sliding window in the adaptive sliding window method is adjusted based on the average time interval of user behavior.

4. The method for generating a dynamic advertising delivery strategy based on user behavior according to claim 1, characterized in that: After determining the association relationship, the following steps are also included: Differential privacy technology is used to perturb the association rules corresponding to the association relationship to prevent the leakage of sensitive attribute information.

5. The method for generating a dynamic advertising delivery strategy based on user behavior according to claim 4, characterized in that: Using differential privacy technology to perturb the association rules corresponding to the association relationship to prevent the leakage of sensitive attribute information includes the following steps: B1. Determine the set of association rules to be perturbed and assign an initial privacy budget to each association rule; the sum of the initial privacy budgets must satisfy the preset total privacy budget constraint. B2. Evaluate the sensitivity of each association rule by considering the sensitive attribute types involved, the support of the association rule, and the confidence of the association rule; B3. Based on the sensitivity of the association rules, adjust the privacy budget allocation for each association rule; association rules with higher sensitivity are allocated smaller privacy budgets, while association rules with lower sensitivity are allocated larger privacy budgets; B4. Based on the adjusted privacy budget and the data type of the association rule, select a differential privacy mechanism for each association rule. Differential privacy mechanisms include the Laplace mechanism and the Gaussian mechanism. B5. Use the selected differential privacy mechanism to perturb the association rules and generate perturbed association rules; B6. Verify the perturbed association rules to ensure that they meet the preset support constraints, confidence constraints, and privacy protection constraints. If the constraints are not met, return to step B3 and readjust the privacy budget allocation.

6. The method for generating a dynamic advertising delivery strategy based on user behavior according to claim 5, characterized in that: The specific steps in step B2 include: Construct a sensitive attribute type grading system, which divides different sensitive attribute types into multiple levels and assigns a corresponding grade score to each level; According to the support of the association rule, the number of users covered by the rule is calculated, and the support score is determined based on the number of users; the more users covered, the higher the support score; Based on historical association rule data, the confidence of the association rule is evaluated to obtain a confidence score; The sensitivity score of the association rule is obtained by taking a weighted sum of the grade score, support score and confidence score; the weights of the support score and confidence score are both lower than the weight of the grade score.

7. A system for generating a dynamic advertising delivery strategy based on user behavior using the method for generating a dynamic advertising delivery strategy based on user behavior according to any one of claims 1 to 6, characterized in that: include: A construction module is used to construct a user behavior sequence by acquiring user behavior data and associated context information; User behavior sequence is used to record user behavior trajectory; The recognition module is used to analyze user behavior sequences and contextual information, identify behavior patterns and contextual features that are statistically associated with sensitive attribute groups, and determine the correlation between behavior patterns and sensitive attribute groups; The association relationship is used to dynamically reflect the association strength between the behavior pattern and the sensitive attribute group; The evaluation module is used to assess the probability of different sensitive attribute groups receiving specific ads based on the correlation between behavioral patterns and sensitive attribute groups and the current advertising delivery strategy. It also calculates the difference in ad reception probabilities between groups to derive a quantitative disparity risk metric. This disparity risk metric is used to quantify the impact that advertising delivery strategies may have on the unequal ad reception opportunities for different user groups. An adjustment module is used to use the difference risk measurement as a constraint condition to adjust the advertising rules to meet the preset difference risk limit and obtain an adjusted advertising strategy; The control module is used to deliver advertisements according to the adjusted advertisement delivery strategy.

Images