Potential advertisement audience identification method

Abstract

The invention relates to a potential advertisement audience identification method. Comprising the following steps of: marking time weights on all behaviors of a user by adopting a behavior delay window concept, introducing behavior silence period logic, and identifying the user which does not clearly express an intention but has a behavior sequence in an abnormal set; a potential user state classifier is constructed, and users are divided into dominant intentions, potential intentions and irrelevant users; constructing a multi-time-scale behavior flow graph by using a sliding time window, and respectively calculating a short-term impulse behavior track and a long-term habitual behavior track; performing intention prediction on the behavior flow through an intention evolution modeling network, and outputting advertisement preference trend points which are not dominant but tend to be stable; adopting a multi-dimensional interest interference discrimination mechanism and real intention clarification; establishing a behavior stability evaluation model, calculating the behavior backtracking coverage degree of a certain point of interest, and excluding one-time behaviors; a cross-modal intention consistency analysis mechanism is introduced; and constructing an interest exclusion list, and filtering unstable interests or heterogeneous behaviors.

Description

Technical Field

[0001] The present invention relates to a method for identifying potential advertising audiences. Background Art

[0002] Currently, traditional advertising audience identification mainly relies on users' explicit behavior data, such as clicks, searches, browsing, purchases, etc. Although these methods can help the advertising system identify potential audiences to a certain extent, there are still many deficiencies and drawbacks. First, traditional methods rely too much on explicit behaviors. When users do not perform these explicit behaviors, the advertising system cannot effectively identify their potential advertising interests. For example, a user may already be interested in a certain product, but because they are busy with other matters and do not click on the relevant advertisement or search for the product, the system cannot capture this potential interest trend, resulting in the missed opportunity for advertising placement. Second, traditional methods usually analyze based on users' immediate behaviors and characteristics, ignoring the evolution process of users' behaviors and the accumulation of potential interests. Most advertising systems often only start intervening in advertising placement when users show clear interests, and this approach may miss some potential audience groups that have not expressed clear intentions, thus greatly reducing the effectiveness of advertising placement.

[0003] Furthermore, traditional methods for identifying potential advertising audiences mainly rely on static label classification based on behavioral characteristics. The system only classifies users into a certain interest category according to their historical behaviors or demographic data. However, with the rapid change of users' interests, this static label classification method fails to capture the migration of users' interests in a timely manner. For example, a user may be highly interested in travel products for a certain period of time, but over time, the user's interest gradually shifts to home decoration or other consumption fields. Traditional methods cannot accurately predict such dynamic interest changes, resulting in the inability of advertisements to be adjusted in a timely manner. More seriously, traditional methods often ignore cross-category interest transfers, that is, users may show interest in a certain field, and this interest may quickly transfer to an entirely different field. Advertising systems usually segment users based on a single category or interest label, making it difficult to identify users who are undergoing interest migration, leading to a significant reduction in the accuracy of advertisements. In addition, traditional methods also face difficulties in identifying cold-start users, that is, new users or users who visit for the first time. Due to the lack of sufficient historical behavior data, the advertising system cannot identify their potential advertising needs in a timely manner. Such users are often excluded from the target of precise advertising due to the lack of sufficient characteristic information, resulting in the loss of advertising opportunities. Furthermore, traditional methods for identifying potential advertising audiences often lack an effective distinction between long-term trends and short-term impulsive behaviors in users' behaviors. This method usually treats all users' behaviors equally, ignoring the difference between users' short-term impulsive behaviors (such as temporary browsing or searching) and long-term stable interests. In fact, many potential advertising audiences' impulsive behaviors in the short term do not represent their long-term interests, and if these short-term behaviors are over-interpreted as potential advertising interests, it may lead to unnecessary advertising placements and waste of advertising budgets. Secondly, existing methods usually analyze advertising interests only through single-dimensional user data (such as clicks, browsing, purchases, etc.), while ignoring the comprehensive impact of multi-dimensional data (such as time, device, location, etc.). In fact, users' advertising interests are not only affected by their behaviors themselves, but also by the context. For example, users may be more interested in certain types of advertisements at a specific time, location, or when using a specific device. Without considering these multi-dimensional information, traditional methods are difficult to accurately capture users' advertising interests in different environments, resulting in a great impact on the accuracy and effectiveness of advertising placements. Existing methods for identifying potential advertising audiences also face the problem of over-reliance on behavioral data. Many advertising systems focus too much on users' behavioral data, believing that all behavioral data is useful, but in fact, not every piece of users' behavioral data can effectively reflect their advertising interests. For example, a user may browse pages related to advertisements frequently during a certain period, but this does not mean that they have the intention to purchase at that time. Many times, users may just browse temporarily or search only to obtain information.Therefore, if the judgment of advertising interest is based solely on single behavioral data, misjudgment and omission are likely to occur, leading to the inefficient operation of the advertising system. Summary of the Invention

[0004] The object of the present invention is to provide a method for identifying potential advertising audiences, so as to solve some of the drawbacks and deficiencies pointed out in the background art.

[0005] The present invention adopts the following technical solutions to solve its above-mentioned technical problems: A method for identifying potential advertising audiences, including the following implementation steps:

[0006] S1. Adopt a potential advertising audience feature recognition mechanism:

[0007] S1.1. Adopt the concept of a behavior delay window, assign time weights to all user behaviors, and introduce the logic of a behavior silence period to identify users whose intention has not been clearly expressed but whose behavior sequences are abnormally concentrated;

[0008] S1.2. Use the behavior pattern change rate including the interest point switching frequency and the access category deviation as the core input signal to identify critical-state users; and construct a potential user state classifier to divide users into three categories: explicit intention, potential intention, and irrelevant users;

[0009] S2. Adopt a time-series preference trajectory modeling and implicit intention reasoning mechanism:

[0010] S2.1. Use a sliding time window to construct behavior flow graphs of multiple time scales, and calculate respectively: short-term impulsive behavior trajectories and long-term habitual behavior trajectories;

[0011] S2.2. Introduce an interest aggregation effect function to identify the gradually focused direction in the trajectory, indicating the subconscious interest convergence point of the user; predict the intention of the behavior flow through an intention evolution modeling network, and output the advertising preference trend points that are not yet explicit but tend to be stable;

[0012] S3. Adopt a multi-dimensional interest interference discrimination mechanism and real intention clarification:

[0013] S3.1. Establish a behavior stability evaluation model, calculate the behavior backtracking coverage of a certain interest point, and exclude one-time behaviors;

[0014] S3.2. Introduce a cross-modal Figure 1 consistency analysis mechanism: including the appearance of relevant content when the user is searching in text, browsing images, and watching videos at the same time, enhancing the intention confidence; otherwise, reducing the weight;

[0015] S3.3. Construct an interest exclusion list to filter unstable interests or heterogeneous behaviors;

[0016] S4. Bind each true intention point to the user context data to form a dynamic intention label for context-weighted intention expression; the label includes interest content, as well as trigger probability time period, location, and device elements; perform spatio-temporal perception-based advertisement matching based on the label, so that the advertisement is only delivered in scenarios with high matching degree, improving the conversion rate.

[0017] Further, the method for constructing the potential advertisement audience feature recognition mechanism includes:

[0018] When identifying potential advertisement audiences, there is a characteristic of the aggregated time distribution of browsing behaviors showing potential: that is, it appears in the latency period where interests have not been clearly expressed, belonging to delayed intention expression behaviors; introduce a behavior delay window model to evaluate the influence degree of each historical behavior on the current interest state within the window; the mathematical expression is:

[0019]

[0020] Where:

[0021] W i represents the time weight value of the user's i-th behavior for the current advertisement interest recognition; t i represents the occurrence timestamp of behavior B i ; T now is the current system time; T w is the set maximum behavior delay window length; α is the time sensitivity coefficient, controlling the strength of the time influence; γ is the non-linear adjustment factor of the decreasing curve, controlling the steepness of the weight function.

[0022] Further, the method for constructing the potential advertisement audience feature recognition mechanism includes:

[0023] Potential advertisement interests evolve from certain stable behaviors; when the user starts to show cross-domain behavior transfers from a stable interest trajectory, including gradually transitioning from travel-related content to decoration, baby products, and car categories, it is predicted that the user is in an interest critical state;

[0024] Define a combined index based on the category jump amplitude and switching frequency to calculate the interest switching intensity in the behavior trajectory and determine whether the user is evolving towards a new advertisement interest area; the change rate function is defined as follows:

[0025]

[0026] Where:

[0027] Δ u represents the interest pattern change rate in the user's overall behavior trajectory; C k represents the content category coding value of the user's l-th behavior, represented by a discrete semantic category ID; |C k+1-C k represents the content category span between two consecutive behaviors; F k is the access frequency of the user in the new category after the behavior switch; is to process the category jump amplitude with the cube root; ln(1 + F k ) is to take the logarithm of the frequency.

[0028] Furthermore, the method for constructing the potential advertising audience characteristic recognition mechanism includes:

[0029] Integrating the aggregation degree of behavior delay weights and the change amplitude of behavior patterns, adopting a three-state user recognition model; introducing a third transitional role: potential intention users, to capture the critical population that has not been explicitly expressed but is very likely to respond to advertisements; designing a probability classification mapping function as follows:

[0030]

[0031] Where:

[0032] Ψ(W, Δ u ) is the classification mapping function representing user state recognition; W = {W1, W2,..., W n} is the weight set of all behaviors of a certain user within the delay window; is the average value of the weights, representing the aggregation degree of the user's recent behaviors; Δ u is the change intensity of the user's interest trajectory; μ1, μ2, μ3 are the weight coefficients corresponding to explicit intention, potential intention, and irrelevant users respectively, satisfying μ1 + μ2 + μ3 = 1, to adjust the recognition bias of different systems; the output result P is a three-dimensional vector, and each component represents the probability that the user belongs to three audience states at the current moment.

[0033] Furthermore, the expression of the weight coefficients corresponding to explicit intention, potential intention, and irrelevant users:

[0034] If the first item is the largest → mark as an explicit intention user;

[0035] If the second item is the largest → mark as a potential intention user;

[0036] If the third item is the largest → identify as the current irrelevant advertising user and do not place advertisements temporarily.

[0037] Furthermore, the method for constructing the time sequence preference trajectory modeling and implicit intention reasoning mechanism:

[0038] Dynamically process the historical behavior data of users through a sliding time window mechanism, and segment the user behavior into subsequences of different time scales according to time continuity, including short-term behavior sequences and medium- and long-term behavior sequences, which are used to capture the immediate interests and stable preferences of users respectively; among them, the short-term window is used to perceive the impulsive behaviors temporarily triggered by users, while the medium- and long-term window is used to observe the persistent tendency of periodic interests or lifestyles; use a time focus weight function to weight and score the user behaviors within each behavior segment, and the calculation method is as follows:

[0039]

[0040] Where:

[0041] Θ(t) is the time focus function, indicating whether the behavior of the user before the current time point t shows an aggregation trend; t0 is the starting time point of the current time window; t is the current evaluation moment; ω(s) is the behavior activity function, indicating the behavior heat of the user at time point s; δ(s) is the behavior intensity function, indicating the response intensity or value density of the user behavior at time point s.

[0042] Furthermore, the method for constructing the time-series preference trajectory modeling and implicit intention reasoning mechanism:

[0043] Construct a user behavior flow graph within multiple time windows, analyze the content association between behavior nodes, and identify whether the behavior gradually focuses on a certain theme; when it is found that the user accesses content clusters with high semantic relevance multiple times at different time scales, it is judged that there is an interest aggregation trend, that is, the behavior path converges in the semantic space, manifested as the formation of subconscious advertising interests; introduce an interest convergence function to quantify this focusing trend as follows:

[0044]

[0045] Where:

[0046] Φ(u) is the interest aggregation degree score of user u, and the higher it is, the more the user behavior continuously approaches a certain type of content; n is the number of aggregation themes, that is, the number of potential interest directions identified in the current behavior trajectory; κ i is the content semantic similarity score between the i-th theme direction and the user's historical behavior; ζ i is the time concentration degree of the user for this theme, reflecting the persistence and density on the behavior theme; ρ is the non-linear adjustment index, which is used to enhance the sensitivity of the model to the focus of high-density content.

[0047] Furthermore, the method for constructing the time-series preference trajectory modeling and implicit intention reasoning mechanism:

[0048] Combining multi-time scale behavior data with the interest aggregation index, a user intention evolution model is established to predict the future advertising interest trend of users within a certain period of time; to determine whether the interest is evolving into a certain type of advertising audience tendency; and to depict the trend, the following interest evolution prediction function is used:

[0049]

[0050] Where:

[0051] Ξ(u,t) is the prediction function, which is used to evaluate the potential score of user u evolving into an advertising audience at the future time point t+ε; ε is the prediction time step, approaching infinitesimal, indicating fine-grained time prediction; Φ(u) is the interest aggregation index at the current moment, serving as the trend basis; Γ(u,s) is the sensitivity function of the user to content or advertising type at time s, obtained by combining historical behavior and context.

[0052] The potential advertising audience identification method of the present invention has significant beneficial effects and can achieve the following core advantages in the fields of advertising placement and user interest identification:

[0053] Through multi-level data analysis and model construction, it is possible to capture the possible advertising interests of users in advance when they have not yet clearly expressed their advertising intentions. This ability to identify potential advertising audiences in advance helps the advertising system to intervene when the earliest signs of interest appear, avoiding the limitations of traditional methods that only rely on explicit user behaviors (such as clicks and purchases); by combining multi-time scale behavior data and the interest aggregation index.

[0054] Using the interest evolution prediction function and combining the time step to predict the evolution trend of users' future interests. This method can track the subtle changes in users' interests and predict the advertising interests that users may generate in a certain future period, thus realizing forward-looking advertising placement. This time-evolution-based prediction enhances the forward-looking and accuracy of advertising, providing more intelligent advertising placement decision support for advertisers;

[0055] By introducing the new concept of potential intention users, those users who have not explicitly expressed their advertising interests but have strong potential interests are identified, avoiding the ineffective push of "irrelevant advertising users" in traditional advertising placement. By accurately judging the interest trend of users, the system can reduce the waste of advertising and improve the utilization efficiency of advertising budgets; methods such as the behavior delay window model and the interest pattern change rate are adopted, which can flexibly adapt to the behavior characteristics of different users and accurately capture the impulse behaviors or long-term stable preferences of users in the short term; through this accurate and personalized advertising placement method, users will see more relevant and demand-compliant advertising content. Description of the Drawings

[0056] Figure 1 This is the flowchart of the method for identifying potential advertising audiences of the present invention.

[0057] Figure 2 This is the flowchart of the method for constructing the mechanism for identifying the characteristics of potential advertising audiences of the present invention.

[0058] Figure 3 This is the flowchart of the method for constructing the mechanism for modeling the time - sequence preference trajectory and inferring implicit intentions of the present invention. Detailed implementation manners

[0059] The following makes a detailed description of the specific implementation manners of the present invention with reference to the accompanying drawings.

[0060] Refer to Figure 1 , for the method of identifying potential advertising audiences of the present invention, through in - depth analysis of user behavior, potential audiences who have not clearly expressed their interest in advertisements are identified. It includes adopting a mechanism for identifying the characteristics of potential advertising audiences, which mainly consists of two key parts. First, the concept of a behavioral delay window is introduced to identify user behaviors whose intentions have not been clearly expressed. The purpose of the behavioral delay window is to solve the phenomenon that potential advertising interest is not always manifested by explicit behaviors (such as clicks, searches) in the short term. Therefore, by timestamping all user behaviors and assigning a time weight to each behavior, the system can identify users who exhibit frequent behaviors within a specific time period. These users often do not clearly express their interest, but their behavior sequences show a certain concentrated trend. The behavioral silence - period logic further enhances the effectiveness of this mechanism. By setting a time window, it is considered that even if some behaviors do not immediately express advertising intentions, potential interest still accumulates over time within the silence period. This stage is particularly important because users have already generated potential interest in certain advertising categories or product areas, but have not expressed it through explicit behaviors (such as clicking on an advertisement or searching for specific content). Next, the behavioral pattern change rate and the interest - point switching frequency are used as the core input signals for identifying critical - state users. The behavioral pattern change rate is used to measure whether there has been a significant deviation in user behavior, and this change usually indicates a shift in user interest. For example, when a user switches from regular social - media browsing to starting to pay attention to the detailed information of a specific product, this indicates that the user is undergoing a potential interest - conversion process.

[0061] In addition, the point-of-interest switching frequency and access category offset can help the system identify users with a strong potential interest conversion trend. Although these users have not expressed a clear interest in advertisements, their behavioral trajectories have shown a high degree of variability, indicating that they are about to cross the boundary from "latent interest" to "explicit interest". To further improve the accuracy of identifying potential advertising audiences, the system classifies users into three categories: explicit-intention users, latent-intention users, and irrelevant users. This classification is established based on the dynamic evolution of users' behaviors and interests. Explicit-intention users refer to those who have expressed a clear interest through explicit behaviors such as clicking and purchasing. Latent-intention users are those who, although they have not shown explicit interest, their behavioral trajectories indicate that they are evolving into advertising audiences. Irrelevant users are those who do not show any trend of advertising interest. To achieve this classification, the system constructs a potential user status classifier, which will classify users into one of the above three categories according to factors such as the user's behavior pattern, change rate, and degree of behavioral focus, so as to adopt different strategies in advertising delivery.

[0062] The second part focuses on the modeling of temporal preference trajectories and the implicit intention reasoning mechanism. This mechanism aims to identify users' potential advertising interests by analyzing the evolution process of their behaviors, even if these interests have not been expressed through explicit behaviors (such as clicking or searching). First, through the sliding time window mechanism, the system divides the user's behavior data according to time to form different time windows, and the behavior data within each window will be regarded as a subsequence. The system analyzes these behavior data to construct multi-time-scale behavior flow graphs. The role of these behavior flow graphs is to identify the user's behavioral trajectories at different time scales, especially short-term impulsive behavioral trajectories and long-term habitual behavioral trajectories. Short-term impulsive behavioral trajectories usually manifest as quick decisions made by users in a short period of time, such as temporary purchases and quick browsing, reflecting the user's momentary interests. Long-term habitual behavioral trajectories, on the other hand, reflect the user's stable interests and preferences over a long period of time, such as when a user browses a certain type of product for a long time or continuously visits a certain website. These trajectories reflect the user's basic needs and preferences. By comprehensively analyzing these short-term and long-term trajectories, the system can more accurately identify the changes in the user's interests at different time scales, thereby further improving the prediction accuracy of advertising interests.

[0063] Next, to further understand the user's potential advertising interests, the system introduces an interest aggregation effect function, which is used to identify the gradually focused interest points in the behavior trajectory. Specifically, when the user's behavior trajectory shows frequent access to a specific field or product within multiple time windows, the system will identify this trend and evaluate whether the user is forming a subconscious interest convergence point for that field or product. The interest aggregation effect function helps the system identify potential interests that have not been explicitly expressed by calculating the degree of focus in the user's behavior. Finally, combining the above-mentioned behavior flow graph and the interest aggregation effect, the system introduces an intention evolution modeling network. The role of this network is to simulate the evolution process of the user's interests and predict the user's future interest trends based on the user's behavior data. By modeling the behavior flow, the system can output the trend points of advertising preferences that are not yet explicit but tend to be stable, predicting the advertising interests that the user will show in the future. This prediction not only depends on the current behavior data but also comprehensively considers the historical change trends of the user's interests, so as to be able to identify the user's advertising preferences in advance before the user clearly expresses their interests.

[0064] The third part focuses on the multi-dimensional interest interference discrimination mechanism and the clarification of the true intention. Its core purpose is to improve the judgment accuracy of the true intention by effectively identifying and eliminating those behaviors that do not truly reflect the user's advertising interests. In this process, first, a behavior stability evaluation model is established to solve the problem of one-time behavior interference in the user's behavior. These behaviors often cannot reflect the user's long-term interests or potential advertising preferences. For example, some users have made a purchase or search only because of a special event, but this behavior does not represent their continuous attention to a certain interest point or potential advertising demand. Therefore, the system judges the user's long-term attention degree to an interest point by calculating the behavior backtracking coverage of that interest point.

[0065] Specifically, the behavior backtracking coverage evaluates the breadth and depth of the user's past behaviors at a specific interest point. The more frequent and continuous the behaviors are, the greater the attraction of that interest point to the user. Incidental single behaviors will be regarded as one-time behaviors and thus excluded from the targets of advertising placement. Next, the system introduces cross-modal meaning Figure 1Consistency analysis mechanism, which enhances the confidence of intent by analyzing the behavioral consistency of users across different modalities. In today's Internet environment, users' interests are usually cross-platform and cross-modal. For example, in scenarios such as text search, image browsing, and video watching, users may show potential interest in a certain advertisement content. Therefore, the system analyzes whether users are interested in relevant content simultaneously across different modalities. If users show consistent interest behavior across multiple modalities, it can be considered that the user has a stronger intent for this interest point, and the system will enhance the confidence of this intent; conversely, if the user's behavior is inconsistent across different modalities, the recognition of this intent will be downgraded to avoid misjudgment. After the above steps are completed, the system further filters out unstable interests or irrelevant behaviors by constructing an interest exclusion list. Through this list, the system can effectively exclude users' interests with large fluctuations, instability, or cross-category, as these users are not real potential advertising audiences. For example, if a user frequently switches interest categories within a short period of time, or shows occasional interest in certain content without continuous attention, these will be regarded as unstable interest points by the system and thus excluded from the precise advertising audience. Overall, by combining behavioral stability, cross-modal consistency analysis, and the interest exclusion mechanism, the system can effectively identify and exclude user behaviors with low relevance or instability to advertisements, improving the accuracy of identifying real advertising intent.

[0066] The fourth part focuses on the construction of dynamic intent tags and spatio-temporal aware advertising matching. The key goal of this part is to generate advertising placement tags with spatio-temporal awareness based on users' interest and behavior data, combined with their context information, so as to ensure that advertisements are only pushed in the most appropriate scenarios to improve the conversion rate of advertisements. First, the system binds each real intent point to the user's context data. This process means comprehensively analyzing and combining the user's interest behavior with the specific situation they are in (such as time, location, device, etc.). For example, a user browses information about a certain electronic product through their mobile phone at night. This behavior is not only an interest in the product itself but also restricted by specific context factors such as the night time period and the type of device used. Therefore, the system will conduct a comprehensive analysis of each real intent point, bind these interest behaviors to the user's context data, and form a comprehensive expression of the interest point.

[0067] Next, the system generates dynamic intent tags for context-weighted intent expression through this binding. These tags not only contain the user's interested content but also elements such as the triggering probability period, location, and device where the user's behavior occurs. For example, if a user browses an advertisement for fitness equipment through a desktop computer at noon, then the user's intent tag not only contains the interested content of fitness equipment but also contains time information (noon), device information (desktop computer), and location information (such as home or office). These elements together determine the best timing and scenario for advertising push. In this way, the user's interest is refined into an information tag with multiple dimensions, further improving the accuracy of advertisement matching. Based on these dynamic intent tags, the system will perform spatio-temporal perception-based advertisement matching, that is, the system intelligently selects the most suitable advertisement scenario for push according to information such as the period, location, and device in the tags. For example, if a user browses an advertisement related to breakfast food through a mobile phone in the morning, the system will, based on this period and device information, preferentially select to push breakfast-related advertisements in the morning and through mobile devices, rather than pushing advertisements at irrelevant times or on irrelevant devices. Through this precise advertisement delivery mechanism, advertisements can not only reach users at the moment when they are most responsive but also improve the relevance and attractiveness of advertisements according to context data, thereby enhancing the conversion rate of advertisements.

[0068] Embodiment 1:

[0069] Combined with the attached Figure 2 , in this embodiment, an e-commerce platform designs a potential advertisement audience recognition system, and the goal is to identify in advance those potential customers who have the intention to purchase a certain electronic product based on the user's behavior data. This process requires the use of potential advertisement audience feature recognition. User Xiaoming has visited web pages related to smart watches in multiple scenarios recently, but has not explicitly clicked the purchase button. Xiaoming's behavior shows an aggregation of potential browsing behavior, that is, although he has not explicitly expressed the intention to purchase (such as not searching for "buy smart watch" or clicking on the advertisement), he has visited relevant web pages multiple times, and the system believes that his interest is gradually forming subconsciously. To identify this potential interest, it is necessary to evaluate the impact of these behaviors on the current advertisement interest through a behavioral delay window model.

[0070] The system will use the following mathematical formula to evaluate the impact of each behavior on Xiaoming's advertisement interest:

[0071]

[0072] Where: W i is the weight value of Xiaoming's i-th behavior on the current advertisement interest; t i is the timestamp of Xiaoming's i-th visit to the smart watch page; T now is the current system time; T wis the maximum behavioral delay window length, set to 72 hours (3 days), which means only behaviors within the past 3 days are considered; α is the time sensitivity coefficient, set to 0.5, controlling that the closer the behavior is to the current moment, the greater the impact; γ is the non-linear adjustment factor of the decreasing curve, set to 2, controlling the steepness of the impact of time on the weight.

[0073] Suppose Xiaoming's first behavior t1 is to access the smartwatch page at 12:00 noon 3 days ago, and the current time is 12:00 noon on a certain day (T now = 12:00), then the time stamp t1 of his first behavior is 72 hours different from the current system time. Calculated according to the above formula:

[0074]

[0075] Xiaoming's second visit t2 occurred 48 hours ago, and the time difference is 48 hours. Therefore:

[0076]

[0077] Xiaoming's third visit t3 occurred 24 hours ago, and the time difference is 24 hours:

[0078]

[0079] In this way, the system calculates the weight of each behavior of Xiaoming in terms of time. The closer the behavior is to the current time, the greater the weight. Suppose Xiaoming has visited the smartwatch web page three times within the past 72 hours, and the calculated time weight values are W1 = 0.5, W2 = 1, W3 = 1 respectively.

[0080] By calculating the weight values of all of Xiaoming's behaviors, the potential interest value of Xiaoming in the smartwatch advertisement is obtained. These interest points will be assigned different weights, which in turn affect the advertisement delivery strategy. For example, if the system finds that most of the weight values of the behaviors are relatively high (such as W2 and W3 are both close to 1) when calculating the total weight of these behaviors, it means that Xiaoming's interest in the product has been gradually accumulating during this period. The system will be more confident in judging that he is a potential advertisement audience, even though he has not explicitly expressed a purchase intention.

[0081] Suppose after analyzing Xiaoming's interests, the system delivers the advertisement to the time period when he is more likely to click (for example, Xiaoming is more inclined to browse products between 8 pm and 10 pm) and the device he usually uses (such as a mobile phone). At this time, spatio-temporal perception-based advertisement matching can display the advertisement on the most suitable time, location and device according to Xiaoming's behavior pattern. For example, if Xiaoming is in the time period when these high-weight behaviors (W2 and W3) occur, the system will choose to deliver the advertisement of the smartwatch on his mobile phone to maximize the conversion rate of the advertisement.

[0082] Through this potential advertising audience identification method, the system can not only identify Xiaoming's potential interest in smart watches even when he has no explicit purchase behavior, but also accurately predict and deliver advertisements at appropriate times, on appropriate devices, and in appropriate contexts. This approach improves the accuracy and conversion rate of advertising. Especially when users' behaviors have not clearly expressed interest, the system can still effectively predict their advertising needs based on past behavior trajectories and improve the effectiveness of advertisements through intelligent matching. Although Xiaoming has not clicked on any advertisements or searched for relevant content, the system has identified the trend of change in his interests based on his behavior data in the past few days, thus ensuring that the advertisement can reach him at the most appropriate moment and ultimately increasing the purchase conversion rate. User Xiaoming used to frequently browse travel websites, but recently his behavior trajectory has started to show cross-domain changes and he has gradually started to browse pages related to home decoration and baby products.

[0083] Use the rate of change function to quantify the degree of jump in user interest from one domain to another. The definition of the rate of change function is as follows:

[0084]

[0085] Where: Δ u is the rate of change of the interest pattern in Xiaoming's overall behavior trajectory, representing whether he has made a smooth transition from one interest domain to another; C k is the content category code to which Xiaoming's kth behavior belongs, represented by a discrete semantic category ID, such as travel = 1, home decoration = 2, baby products = 3, etc.; |C k+1 - C k | is the content category span between two consecutive behaviors, that is, the span of Xiaoming's behavior from one category (such as travel) to another category (such as home decoration). The larger the value, the greater the conversion of the user's interest; F k is the access frequency of Xiaoming in the new category after the behavior switch, that is, the number of times he views relevant content in the new category. The higher the frequency, the stronger Xiaoming's interest in the new category; Use the cube root to process the category jump amplitude to reduce the impact of extreme changes on the overall result; ln(1 + F k ) taking the logarithm of the frequency is to avoid the excessive influence of some small access frequencies on the result and to emphasize the interest transfer of high frequencies.

[0086] Suppose Xiaoming has carried out the following behaviors in the past month:

[0087] The 1st behavior: He browsed on travel-related websites, belonging to the travel category (category code is 1), and the access frequency was 5 times;

[0088] The second behavior: He began to browse websites related to home decoration (category code: 2), with an access frequency of 10 times;

[0089] The third behavior: He then browsed web pages related to baby products (category code: 3), with an access frequency of 7 times.

[0090] Now, let's calculate the interest switching intensity Δ of Xiaoming u 。

[0091] For the category span |C2 - C1| between the first and second behaviors:

[0092] The category jumps from travel (1) to home decoration (2), so the span is |2 - 1| = 1;

[0093] The access frequency F2 = 10. Therefore, taking the logarithm of the frequency gives ln(1 + 10) = ln(11) ≈ 2.398;

[0094] The cube root of the category span:

[0095] So, the change rate between the first and second behaviors is:

[0096] Δ1 = 1 × 2.398 = 2.398

[0097] For the category span |C3 - C2| between the second and third behaviors:

[0098] The category jumps from home decoration (2) to baby products (3), so the span is |3 - 2| = 1;

[0099] The access frequency F3 = 7. Therefore, taking the logarithm of the frequency gives ln(1 + 7) = ln(8) ≈ 2.079;

[0100] The cube root of the category span:

[0101] So, the change rate between the second and third behaviors is:

[0102] Δ2 = 1 × 2.079 = 2.079

[0103] We can obtain the total interest change rate Δ of Xiaoming u :

[0104] Δ u = Δ1 + Δ2 = 2.398 + 2.079 = 4.477

[0105] The interest pattern change rate Δ of Xiaoming u= 4.477. This value indicates that his interest has a significant jump and a high change rate in switching from one field (travel) to another field (home decoration and baby products) in a short period of time. This shows that he is undergoing a potential interest transformation process, expanding his interest from the travel interest field to advertising-related fields such as home decoration and baby products. Based on this change rate, the system can determine that Xiaoming is in a critical state of interest, that is, he has an advertising demand for categories such as home decoration and baby products, although he has not explicitly expressed a purchase intention at present.

[0106] Based on this calculation, the system can further infer Xiaoming's potential advertising interests and push corresponding advertisements to him in appropriate time and space scenarios. For example, if Xiaoming often uses his mobile phone to browse relevant content at night, the system can accurately push advertisements for home decoration and baby products to him through his mobile phone at night, rather than wasting advertising resources at other irrelevant times and on other devices. Through this precise interest prediction and advertisement matching, the conversion rate of advertisements can be significantly improved.

[0107] Next, through the probability classification mapping function, accurately determine whether Xiaoming belongs to an explicit intention user, a potential intention user, or an irrelevant advertisement user.

[0108] User Xiaoming is an online shopping user. Recently, he has become interested in smart watches and started browsing relevant web pages frequently. However, although Xiaoming has not immediately clicked on the advertisement or purchased a smart watch, his behavior trajectory still shows potential advertising interests. The goal is to determine whether Xiaoming belongs to a potential intention user based on his behavior data, combined with the delay weight and the change in behavior pattern, and adjust the advertisement delivery strategy according to his interest pattern.

[0109] Use the following probability classification mapping function to evaluate Xiaoming's advertising intention:

[0110]

[0111] W = {W1, W2, …, W} represents the weight set of all Xiaoming's behaviors in the past period of time, calculated through the behavior delay window model;

[0112] represents the average weight of these behaviors, reflecting the concentration of Xiaoming's recent behaviors;

[0113] Δ u represents the change intensity of Xiaoming's interest trajectory, measuring whether he is shifting from one interest field to another;

[0114] μ1, μ2, μ3 are the weight coefficients corresponding to explicit intention, potential intention, and irrelevant users respectively, and satisfy μ1 + μ2 + μ3 = 1. These coefficients are used to adjust the recognition bias of different categories of users.

[0115] In the past few days, Xiaoming's behavioral trajectory is as follows:

[0116] The first behavior: He visited the product details page of the smart watch. The behavior occurred 72 hours ago, and the calculated behavior weight W1 = 0.5;

[0117] The second behavior: He visited the web page of the same product 48 hours ago, and the behavior weight W2 = 0.8;

[0118] The third behavior: He visited the web page again 24 hours ago, and the weight W3 = 1.

[0119] Among these behaviors, Xiaoming's recent behaviors (W2 and W3) are relatively more important, reflecting his strong interest in this product in the short term. The average weight of these behaviors is:

[0120]

[0121] It is assumed that Xiaoming's interest trajectory has changed significantly. Originally, he had been focusing on travel-related content, but recently he has started to frequently browse pages related to smart watches, which means that his interest has shifted across fields. The intensity of his interest change Δ u has been calculated through the behavior pattern change rate formula and is set to 4.5 (for the specific calculation process, please refer to the previous interest change rate calculation). This change intensity indicates that his interest in smart watches is gradually increasing.

[0122] Now, the probability classification mapping function can be calculated based on the above data. The following weight coefficients are set:

[0123] μ1 = 0.6 (the proportion of explicit intention users is relatively high);

[0124] μ2 = 0.3 (the proportion of potential intention users is relatively low);

[0125] μ3 = 0.1 (the proportion of irrelevant users is relatively low).

[0126] Then, Xiaoming's state probability vector P is calculated as follows:

[0127] The probability of an explicit intention user:

[0128]

[0129] The probability of a potential intention user:

[0130]

[0131] The probability of an irrelevant user:

[0132]

[0133] According to the calculated probability vector P = [0.0647, 0.2207, 0.0159], we can see that Xiao Ming is a potential user because his potential user probability P2 is the maximum value (0.2207). Therefore, the system will mark Xiao Ming as a potential user, that is, although he has not yet clearly expressed his purchase intention, due to his recent behavior trajectory and interest changes, the system believes that he has a high potential response to smartwatch ads.

[0134] In this way, the advertising system can accurately identify Xiao Ming and push the corresponding advertisement to him at the most appropriate time and space conditions. Assuming that Xiao Ming's device is usually used for browsing between 8 and 10 pm, the system will push smartwatch ads through the phone during these time periods based on Xiao Ming's potential intentions, rather than placing ads at irrelevant time periods or devices. This not only improves the relevance of the ads, but also enhances the conversion rate of the ads.

[0135] Embodiment 2:

[0136] Combined with Figure 3 In this example, Xiao Ming is an active user of the e-commerce platform. Recently, he frequently browsed pages about sports shoes, smart watches and other products, but did not show any clear purchasing behavior. The system dynamically processes Xiao Ming's historical behavior data and tries to identify whether he has potential interest in a certain type of product (such as sports shoes or smart watches), although he has not expressed his interest through explicit behavior (such as clicking on ads or searching for related products).

[0137] Two time windows are set for Xiao Ming: a short-term window and a medium- to long-term window. The short-term window is used to capture Xiao Ming's temporary interests or impulsive behaviors, while the medium- to long-term window is used to observe Xiao Ming's stable interests and long-term tendencies. The short-term window is set to the past 24 hours, while the medium- to long-term window is set to the past week. Within these two time windows, the system will analyze Xiao Ming's behavior and calculate the time focus weight of each behavior.

[0138] In order to identify whether Xiao Ming shows a strong interest in a certain type of product in the short term or medium to long term, the system uses a time focus weight function to weight each piece of behavior data. The calculation formula is as follows:

[0139]

[0140] Among them: Θ(t) is the time focus degree, indicating whether Xiaoming's behaviors before the current time point t show an aggregation trend. If the focus degree is high, it means there are more concentrated behaviors within this time period; t0 is the starting time point of the current time window, set to 24 hours or 7 days before the current evaluation time; t is the current evaluation moment; ω(s) is the behavior activity function, indicating the behavior heat of Xiaoming at a certain moment s, such as access frequency, click times, etc.; δ(s) is the behavior intensity function, indicating the behavior response intensity of Xiaoming, such as residence time, page view depth, etc.

[0141] Set the behavior data of Xiaoming in the past few days:

[0142] Short-term window (within the past 24 hours): Xiaoming browsed the pages related to sports shoes 3 times, staying for 2 minutes, 5 minutes, and 1 minute respectively;

[0143] Medium- and long-term window (within the past week): Xiaoming browsed the pages related to smart watches 2 times, staying for 3 minutes and 4 minutes respectively.

[0144] Set the following parameters:

[0145] ω(s): Behavior activity, scored according to the frequency of page access and page depth. For example, 2 page views indicate an activity of 0.7, and 3 page views are 1;

[0146] δ(s): Behavior intensity, considering the influence of residence time, set the intensity per minute of residence time, and Xiaoming's each residence time directly affects the intensity. For example, the intensity for a 2-minute stay is 0.2, and the intensity for a 5-minute stay is 0.5.

[0147] Within the short-term window, it is set that Xiaoming browsed the sports shoes page three times, staying for 2 minutes, 5 minutes, and 1 minute respectively within the past 24 hours. Calculate the activity and intensity of each behavior, and then obtain the time focus value by weighting:

[0148] The first time browsing the sports shoes page: the activity is 0.7, staying for 2 minutes, and the intensity is 0.2;

[0149] The second time browsing the sports shoes page: the activity is 1, staying for 5 minutes, and the intensity is 0.5;

[0150] The third time browsing the sports shoes page: the activity is 0.7, staying for 1 minute, and the intensity is 0.1.

[0151] According to the formula, calculate the focus degree within this window:

[0152]

[0153] This result indicates that Xiaoming has a relatively high interest focus degree on sports shoes products in the short term.

[0154] In the medium- and long-term window, Xiaoming browsed the smartwatch page twice, staying for 3 minutes and 4 minutes respectively. Assuming the activity level of the behavior within this window is 1 (due to the relatively small number of visits), the intensities are 0.3 and 0.4. Calculated through the same formula, we get:

[0155]

[0156] This indicates that within the past week, Xiaoming's interest focus on smartwatches is relatively low, suggesting that he pays less attention to this category.

[0157] According to the calculation results, Xiaoming shows a relatively high interest focus (1.3875) in the short-term window, while in the medium- and long-term window, his interest is more dispersed and the focus is lower (1.0). Therefore, the system will infer his interest in different categories of advertisements based on these focus results.

[0158] The system sets a threshold: when the focus is greater than 1.2, it is considered that the user has a strong interest in advertisements. In this example, since the focus of Xiaoming's interest in sports shoes in the short-term window is higher than 1.2, the system will target the advertisement to content related to sports shoes, especially during the period when Xiaoming visits again in the future, while the advertisement for smartwatches can be postponed or reduced.

[0159] Through this modeling of temporal preference trajectories and implicit intention reasoning mechanism, the system can accurately identify Xiaoming's potential interest in advertisements, especially when he has not explicitly expressed his interest. This method not only helps the system capture short-term impulsive behaviors but also enables it to identify the potential needs of users in long-term behaviors, thereby providing accurate audience targeting for advertisers and pushing relevant advertisements at the best time.

[0160] Furthermore, through the modeling of temporal preference trajectories and implicit intention reasoning mechanism, construct a user behavior flow graph within multiple time windows to identify whether the user is gradually focusing on a certain theme, especially when the advertisement interest has not been clearly expressed.

[0161] As a user of an e-commerce platform, Xiaoming has recently frequently browsed related products such as smartwatches and sports shoes, but has not shown clear purchase behavior. Nevertheless, Xiaoming's behavior has potential interest, and these behaviors are gradually accumulated within a certain latency period. The system needs to analyze the behavior flow graph of Xiaoming in multiple time windows, identify whether there is a trend of interest aggregation, and infer whether he belongs to the potential advertisement audience.

[0162] To gain a deep understanding of Xiaoming's interest trajectory, two time windows are set for him: a short-term window and a medium- and long-term window. The short-term window is used to capture Xiaoming's temporary interests, which usually reflect impulsive behaviors, while the medium- and long-term window is used to analyze Xiaoming's stable interest trajectory and potential long-term needs. The short-term window is set as the past 24 hours, and the medium- and long-term window is set as the past week. Through these time windows, a behavior flow graph of Xiaoming can be constructed, that is, the behavior nodes within each time period and the content associations between these nodes.

[0163] To determine whether Xiaoming's interests are gradually focusing, an interest aggregation function will be used to quantify the interest convergence trend of him in different time windows. Specifically, when the user's behavior path gradually converges within the semantic space, manifested as the formation of subconscious advertising interests, the system will determine it as interest aggregation. The following is the definition of the interest aggregation function:

[0164]

[0165] Where: Φ(u) is Xiaoming's interest aggregation degree score, and the higher it is, the more he is gradually focusing on a certain type of content; n is the number of aggregation topics, that is, the number of potential interest directions that Xiaoming visits in different time periods (for example, he shows interests in three major categories of products: smart watches, sports shoes, and outdoor activities); κ i is the semantic similarity score between the i-th topic direction and Xiaoming's historical behavior, that is, his attention degree in this interest direction; ζ i is the time concentration, reflecting Xiaoming's behavior persistence and density on this topic; ρ is the non-linear adjustment index, used to enhance the sensitivity to the focus of high-density content.

[0166] It is set that within the past 24 hours, Xiaoming showed strong interest behaviors in the short-term window, mainly concentrated on sports shoes and smart watches. While within the past week, his interests were more scattered, covering sports shoes, smart watches, and outdoor activities. Assign the following data to these behaviors:

[0167] Short-term window behavior (within the past 24 hours):

[0168] Smart watch: Visited 2 times, stayed for 5 minutes each time, and the activity level is 0.8;

[0169] Sports shoes: Visited 3 times, stayed for 4 minutes each time, and the activity level is 1.0;

[0170] Medium- and long-term window behavior (within the past week):

[0171] Smart watch: Visited 5 times, stayed for 6 minutes each time, and the activity level is 0.7;

[0172] Sports shoes: Visited 2 times, stayed for 3 minutes each time, and the activity level is 0.9;

[0173] Outdoor activities: visited 1 time, stayed for 5 minutes, activity level is 0.6.

[0174] Based on these data, first calculate the semantic similarity score (κ i ) and the time concentration (ζ i ) for each topic.

[0175] For sports shoes, Xiaoming has a high activity level and visits multiple times. Therefore, the semantic similarity κ 运动鞋 = 0.9, and the time concentration ζ 运动鞋 = 0.8;

[0176] For smartwatches, although he browsed multiple pages, the overall activity level is slightly lower. The semantic similarity κ 智能手表 = 0.7, and the time concentration ζ 智能手表 = 0.6;

[0177] For outdoor activities, since he only visited once, the semantic similarity κ 户外活动 = 0.5, and the time concentration ζ 户外活动 = 0.4.

[0178] Set the non - linear adjustment index ρ = 1.5, which will enhance the system's sensitivity to high - frequency behaviors (such as visiting the sports shoes page multiple times).

[0179] Based on the above data, calculate the interest aggregation degree Φ(u). It is set that Xiaoming mainly focused on sports shoes and smartwatches in the past 24 hours, and he still maintained a certain degree of attention to smartwatches and sports shoes in the past week. The specific calculation is as follows:

[0180]

[0181] Calculate:

[0182]

[0183] Therefore, the interest aggregation degree is:

[0184] Φ(u)=0.3942 + 0.2926+0.1375 = 0.8243

[0185] Based on the calculated interest aggregation degree Φ(u)=0.8243, it can be concluded that Xiaoming shows a strong interest focus on sports shoes and smartwatches within the short - term window. Although he did not make an immediate purchase, the focus trend of these behaviors indicates that he has the potential to purchase at some point in the future. The system, based on this data, determines that Xiaoming belongs to the potential advertising audience and decides to push advertising content related to sports shoes and smartwatches under appropriate space - time conditions in the future.

[0186] The system sets a focus threshold of 0.7. When the interest aggregation degree is greater than this value, the system will display advertisements. Therefore, Xiaoming's focus meets the criteria for advertisement display, and the system will push relevant advertisements during the periods when Xiaoming is active (such as from 8 pm to 10 pm, browsing through mobile phones).

[0187] Continue to predict Zhang Wei's future advertisement interest trend by modeling his historical behavior, combining his interest aggregation index (Φ(u)) and sensitivity to content (Γ(u, s)).

[0188] Zhang Wei is an active online shopping user. Recently, he has started to frequently browse products related to smart watches and sports shoes, although he has not explicitly expressed his interest by clicking on advertisements or making purchases. The goal of the system is to predict whether he will be interested in a certain type of advertisement in the future by using Zhang Wei's historical behavior data (including short-term and medium- to long-term browsing behaviors), combined with his interest aggregation index, and to provide advertisers with accurate timing for advertisement placement. For this purpose, a user intention evolution model needs to be established, which can predict Zhang Wei's advertisement interest trend in the future for a period of time.

[0189] To accurately predict Zhang Wei's future advertisement interest, the following interest evolution prediction function is used:

[0190]

[0191] Among them, Ξ(u, t) is the prediction function, representing the advertisement interest trend score of Zhang Wei at the future time point t + ε;

[0192] ε is the time step, approaching infinitesimal, meaning fine-grained time prediction;

[0193] Φ(u) is Zhang Wei's interest aggregation index at the current moment, measuring his current potential interest in a certain type of advertisement content;

[0194] Γ(u, s) is Zhang Wei's sensitivity function to advertisement content or type at time point s, representing his interest intensity and response ability to advertisements.

[0195] It is set that the following information is extracted from Zhang Wei's behavior data:

[0196] In the past 24 hours, Zhang Wei has shown a high interest in smart watches and sports shoes. The system calculates that the interest aggregation index Φ(u) = 0.85, indicating that Zhang Wei's interest is gradually focusing on these two types of products;

[0197] The system further analyzed Zhang Wei's browsing history in the past week and obtained his sensitivity function Γ(u, s). It is set that within the past 24 hours, Zhang Wei has a high sensitivity to smartwatch ads (Γ(u, s) = 0.9) and a low sensitivity to sports shoe ads (Γ(u, s) = 0.6);

[0198] The time step ε is set to 1 hour, that is, a prediction is made every hour.

[0199] Set the current evaluation time point t to the current moment. It is necessary to predict the evolution trend of Zhang Wei's advertising interest within the next 1 hour (i.e., t + 1 hour). Based on the existing data, the interest aggregation degree Φ(u) and the sensitivity function Γ(u, s) will be used for calculation.

[0200] For smartwatch ads, set its sensitivity within the next 1 hour to 0.9, and the calculation is as follows:

[0201]

[0202] Since ε = 1 hour, it can be simplified to:

[0203] Ξ 智能手表 = 0.85 · 0.9 = 0.765

[0204] For sports shoe ads, set its sensitivity within the next 1 hour to 0.6, and the calculation is as follows:

[0205]

[0206] Similarly, it is simplified to:

[0207] Ξ = 0.85 · 0.6 = 0.51

[0208] Sports shoes

[0209] Through calculation, the evolution trend scores of Zhang Wei's interest in smartwatch and sports shoe ads within the next 1 hour are obtained:

[0210] The interest score for smartwatch ads is 0.765, indicating that he has a strong potential interest in smartwatch ads;

[0211] The interest score for sports shoe ads is 0.51, indicating that his interest in sports shoe ads is relatively low.

[0212] Based on this result, the advertising system determines that Zhang Wei has a relatively high interest in smartwatch advertisements within the next hour. Therefore, it will prioritize pushing relevant smartwatch advertisement content, while the sports shoe advertisements can be postponed or the frequency of their delivery can be reduced. In this way, the advertising system can infer his potential advertising needs through historical behavior data and interest prediction models even when the user has not explicitly expressed interest.

[0213] Through the interest evolution prediction function, the system successfully predicted Zhang Wei's potential interest in advertisements within the next hour. This prediction is not only based on the current behavior data but also combines Zhang Wei's historical responses to advertisements and his current interest focus, thus achieving more precise advertisement delivery. Advertisers can make dynamic advertisement placement decisions based on the trend of Zhang Wei's interest changes in the future time period to maximize the conversion rate and effectiveness of advertisements. Through this time-based predictive advertisement placement strategy, the advertising system not only improves the relevance and timeliness of advertisements but also ensures that advertisements are only pushed in the most relevant spatio-temporal scenarios, thereby enhancing the user experience and the return on investment for advertisers.

Claims

1. A method for identifying potential advertising audiences, characterized in that Including the following steps: S1. Adopt a potential advertising audience feature recognition mechanism: S1.

1. Adopt the concept of a behavior delay window, assign time weights to all user behaviors, and introduce the logic of a behavior silence period to identify users whose intention has not been clearly expressed but whose behavior sequences are abnormally concentrated; S1.

2. Use the behavior pattern change rate including the frequency of point-of-interest switching and the deviation of access categories as the core input signal to identify critical-state users; and construct a potential user state classifier to divide users into three categories: explicit intention, potential intention, and irrelevant users; S2. Adopt a time-series preference trajectory modeling and implicit intention reasoning mechanism: S2.

1. Use a sliding time window to construct a behavior flow graph with multiple time scales, and calculate respectively: short-term impulsive behavior trajectories and long-term habitual behavior trajectories; S2.

2. Introduce an interest aggregation effect function to identify the gradually focused direction in the trajectory, which represents the subconscious interest convergence point of the user; perform intention prediction on the behavior flow through an intention evolution modeling network, and output the advertising preference trend points that are not yet explicit but tend to be stable; S3. Adopt a multi-dimensional interest interference discrimination mechanism and real intention clarification: S3.

1. Establish a behavior stability evaluation model, calculate the behavior backtracking coverage of a certain point of interest, and exclude one-time behaviors; S3.

2. Introduce a cross-modal intention consistency analysis mechanism: including the appearance of relevant content when the user is simultaneously engaged in text search, image browsing, and video viewing to enhance the intention confidence; Otherwise, reduce the weight; S3.

3. Construct an interest exclusion list to filter unstable interests or heterogeneous behaviors; S4. Bind each real intention point to the user context data to form a dynamic intention label with context-weighted intention expression; the label contains interest content, as well as trigger probability time period, location, and device elements; Perform spatio-temporal perception-based advertising matching based on the label, so that the advertisement is only delivered in scenarios with high matching degree, improving the conversion rate.

2. The potential advertising audience identification method according to claim 1, wherein The construction method of the potential advertising audience feature recognition mechanism includes: When identifying potential advertising audiences, it shows the aggregated time distribution characteristics of latent browsing behaviors: that is, it appears in the latency period when the interest has not been clearly expressed, belonging to the behavior of delayed intention expression; introduce a behavior delay window model to evaluate the influence degree of each historical behavior within the window on the current interest state; the mathematical expression is: Where: W i represents the time weight value for identifying the interest of the user's i-th behavior in the current advertisement; t i represents the occurrence timestamp of behavior B i ; T now is the current system time; T w is the set maximum behavior delay window length; α is the time sensitivity coefficient, controlling the strength of the time impact; γ is the non-linear adjustment factor of the decreasing curve, controlling the steepness of the weight function.

3. The potential advertising audience identification method according to claim 2, wherein The construction method of the potential advertising audience feature recognition mechanism includes: potential advertising interests evolve from a certain stable behavior; when a user starts to show cross-domain behavior transfers from a stable interest trajectory, including gradually transitioning from travel-related content to decoration, baby products, and automotive categories, it is predicted that the user is in an interest critical state; define a combined index based on the category jump amplitude and switching frequency to calculate the interest switching intensity in the behavior trajectory and determine whether the user is evolving towards a new advertising interest field.

4. The potential advertising audience identification method according to claim 3, characterized in that The construction method of the potential advertising audience feature recognition mechanism includes: Comprehensively considering the aggregation degree of behavior delay weights and the change amplitude of behavior patterns, adopt a three-state user recognition model; introduce a third transitional role: potential intention users, to capture the critical population that has not been explicitly expressed but is highly likely to respond to advertisements; design a probability classification mapping function as follows: Where: Ψ(W,Δ u ) is a classification mapping function representing user state recognition; W = {W1, W2, …, W n} is the weight set of all behaviors of a certain user within the delay window; is the average value of the weights, representing the aggregation degree of the user's recent behaviors; Δ u is the change intensity of the user's interest trajectory; μ1, μ2, μ3 are the weight coefficients corresponding to the explicit intention, latent intention, and irrelevant users respectively, satisfying μ1 + μ2 + μ3 = 1, and adjusting the recognition bias of different systems; the output result P is a three-dimensional vector, and each component represents the probability that the user belongs to three audience states at the current moment.

5. The potential advertising audience identification method according to claim 4, wherein The expression of the weight coefficients corresponding to the explicit intention, potential intention, and irrelevant users: If the first item is the largest → Marked as an explicit intention user; If the second item is the largest → Marked as a potential intention user; If the third item is the largest → Identified as the current irrelevant advertisement user, and no advertisement is placed temporarily.

6. The method for identifying potential advertising audiences according to claim 1, wherein The method for constructing the time-sequence preference trajectory modeling and implicit intention reasoning mechanism: Dynamically process the historical behavior data of users through the sliding time window mechanism, and segment the user behavior into subsequences of different time scales according to time continuity, including short-term behavior sequences and medium- and long-term behavior sequences, which are used to capture the immediate interests and stable preferences of users respectively; among them, the short-term window is used to perceive the impulsive behaviors temporarily triggered by users, while the medium- and long-term window is used to observe the persistent tendency of periodic interests or lifestyles.

7. The method for identifying potential advertising audiences according to claim 6, wherein The method for constructing the time-sequence preference trajectory modeling and implicit intention reasoning mechanism: Construct a user behavior flow graph within multiple time windows, analyze the content association between behavior nodes, and identify whether the behavior gradually focuses on a certain theme; when it is found that the user accesses highly semantically relevant content clusters multiple times at different time scales, it is judged that there is a trend of interest aggregation, that is, the behavior path converges in the semantic space, manifested as the formation of subconscious advertisement interests.

8. The method for identifying potential advertising audiences according to claim 7, wherein The method for constructing the time-sequence preference trajectory modeling and implicit intention reasoning mechanism: Combine the multi-time scale behavior data and the interest aggregation index to establish a user intention evolution model, which is used to predict the future advertisement interest trend generated by the user within a certain period of time; to judge whether the interest is evolving into a certain type of advertisement audience tendency; and to depict the trend, use the following interest evolution prediction function: Where: Ξ(u,t) is the prediction function, which is used to evaluate the potential score of user u evolving into an advertisement audience at the future time point t+ε; ε is the prediction time step, approaching infinity, indicating fine-grained time prediction; Φ(u) is the interest aggregation index at the current moment, serving as the trend basis; Γ(u,s) is the sensitivity function of the user to the content or advertisement type at time s, obtained by combining historical behavior and context.

Images