A user portrait construction method and system in an internet advertisement recommendation process

Abstract

The present application relates to the technical field of advertisement recommendation, in particular to a user portrait construction method and system in an internet advertisement recommendation process. The method collects terminal acceleration vector, advertisement label, text quantity, sliding trajectory and sliding speed when a user browses an advertisement in real time; obtains an interaction hesitation index based on text quantity, sliding speed size and change and sliding trajectory inflection point curvature; obtains a space stability coefficient based on the direction similarity and amplitude change of the terminal acceleration vector of adjacent time; obtains an immersion attention weight according to the interaction hesitation index and the space stability coefficient; and finally updates an effective interest weight based on the immersion attention weight, historical interest weight and cumulative exposure to construct a user portrait. The present application can effectively strip the physical environment interference, accurately depict the real interaction intention of the user and inhibit the repeated exposure fatigue, and significantly improve the objectivity and accuracy of the user portrait construction.

Description

Technical Field

[0001] This invention relates to the field of advertising recommendation technology, specifically to a method and system for constructing user profiles in the process of internet advertising recommendation. Background Technology

[0002] In the current field of internet advertising recommendation, building accurate user profiles is the core technological foundation for achieving personalized recommendations and intelligent ad delivery. The main purpose of user profiling is to structurally represent multi-dimensional information such as users' historical behavior and interests. By matching these structured tags, ad recommendation engines can push highly relevant ad content to different users, thereby significantly improving ad click-through rates and conversion rates, and optimizing the overall efficiency of ad delivery.

[0003] The current system dynamically updates the interest weights in the user profile based on the user's behavior data when browsing advertisements, thereby capturing changes in user preferences in real time and maintaining the matching degree of recommended content. However, existing behavioral analysis often relies solely on simple statistical indicators such as ad dwell time or clicks, lacking a deep portrayal of users' information reception and behavioral decision-making processes during browsing. For example, a long dwell time might be an ineffective pause due to users leaving their phones idle, rather than a sign of strong interest, thus failing to accurately uncover users' true interaction intentions. Secondly, mobile users often browse ads in highly complex scenarios (e.g., while on public transport, on the subway, or walking), which easily leads to device shaking, causing passive finger tremors or erratic movements. This can easily be misinterpreted as active hesitation or attention, resulting in biased extracted interest features. Finally, users' true interests evolve dynamically and exhibit diminishing marginal utility. Existing interest weight update mechanisms ignore aesthetic fatigue caused by repeated high-frequency exposure of the same type of ad tags, leading to artificially high interest weights for certain frequently exposed tags. This not only easily traps users in a homogenized information cocoon but also severely reduces the objectivity and accuracy of user profiles over long periods.

[0004] Therefore, existing user profile construction methods have obvious limitations in behavioral feature extraction and interest weight update mechanisms. They cannot effectively eliminate physical environment interference and accurately refine users' true interaction intentions, nor can they reasonably suppress interest fatigue caused by repeated exposure. Ultimately, it is difficult to construct objective, accurate and highly dynamic real user profiles. Summary of the Invention

[0005] To address the technical problems of inaccurate user profile construction caused by existing user profiling methods failing to effectively eliminate physical environmental interference and reasonably suppress interest fatigue caused by repeated exposure, the present invention aims to provide a user profile construction method and system in the internet advertising recommendation process. The specific technical solution adopted is as follows: In a first aspect, one embodiment of the present invention provides a method for constructing user profiles during the internet advertising recommendation process, the method comprising the following steps: Real-time data collection of terminal acceleration vector, ad tags, text quantity, scrolling trajectory, and scrolling speed when users browse ads; Based on the number of words in the ad currently being viewed by the user, the magnitude and change of the scrolling rate during the time the user stays on the ad, and the curvature of the inflection point in the scrolling trajectory, the interaction hesitation index of the ad currently being viewed by the user is obtained. Based on the similarity of the direction and the change in the amplitude of the terminal acceleration vector between adjacent moments within the time period during which the user is currently viewing the advertisement, the spatial stability coefficient of the advertisement currently being viewed by the user is obtained. Based on the interaction hesitation index and spatial stability coefficient, obtain the immersive attention weight of the advertisement currently viewed by the user; Based on the immersion attention weight, the historical interest weight corresponding to the ad tag of the ad currently viewed by the user, and the cumulative exposure within a preset historical time period, the effective interest weight of the ad currently viewed by the user is updated to build a user profile.

[0006] Furthermore, the method for obtaining the interaction hesitation index is as follows: The mean and standard deviation of the scrolling rate during the current viewing time of the advertisement are obtained and used as the analysis mean and standard deviation, respectively. Based on the number of words in the advertisement currently being viewed by the user, the mean of the analysis, and the standard deviation of the analysis, the degree of information acceptance of the advertisement currently being viewed by the user is obtained; The sum of the curvatures of all inflection points in the scrolling trajectory during the current viewing time of the advertisement is used as the degree of hesitation of the user regarding the current advertisement. The result of multiplying the degree of information acceptance and the degree of hesitation, and then normalizing the result, is used as the interaction hesitation index of the advertisement currently being viewed by the user.

[0007] Furthermore, the method for obtaining the degree of information acceptance is as follows: The number of characters, the mean of the analysis, and the standard deviation of the analysis are all standardized and mapped to a unified dimensionless interval. The normalized mean and standard deviation of the analysis are added together to determine the sliding stability. The ratio of the standardized number of characters to the degree of scrolling stability is used as the user's acceptance level of the information in the advertisement they are currently viewing.

[0008] Furthermore, the method for obtaining the spatial stability coefficient is as follows: Use the time period during which the currently viewed advertisement is viewed as the target time; When the number of target moments is greater than or equal to the preset number threshold, for any target moment, the result of normalizing the cosine value of the terminal acceleration vector of the target moment and the previous adjacent target moment is used as the directional stability value of the target moment. The average vector of the terminal acceleration vectors at all target times is used as the overall analysis vector; The magnitude of the difference vector between the terminal acceleration vector at the target moment and the overall analysis vector is subtracted to obtain the acceleration change analysis value at the target moment. The ratio of the mean of the directional stability values ​​at all target times to the mean of the acceleration change analysis values ​​is used as the trend stability coefficient. The mean of the magnitudes of the difference vectors between the terminal acceleration vectors of any two adjacent target moments is used as the difference analysis value. The result of normalizing the ratio of the trend stability coefficient to the difference analysis value is used as the spatial stability coefficient of the advertisement currently viewed by the user. When the number of target moments is less than the preset threshold, the preset stability coefficient is used as the spatial stability coefficient of the advertisement currently viewed by the user.

[0009] Furthermore, the method for obtaining the immersion attention weight is as follows: The product of the interaction hesitation index and the spatial stability coefficient is used as the weight of the user's immersive attention to the advertisement currently being viewed.

[0010] Furthermore, the method for obtaining the cumulative exposure is as follows: The total number of times a user views an ad with the same ad tag as the currently viewed ad within a preset historical time period is used as the cumulative exposure.

[0011] Furthermore, the method for obtaining the effective interest weights is as follows: Based on the cumulative exposure, an exposure attenuation factor is obtained; the cumulative exposure and the exposure attenuation factor are negatively correlated. The initial interest weight of the ad currently viewed by the user is obtained by multiplying the immersive attention weight, the preset reference weight, and the exposure decay factor, and then adding the historical interest weight. The minimum value between the initial interest weight and constant 1 is taken as the effective interest weight of the advertisement currently viewed by the user.

[0012] Furthermore, the method for constructing user profiles is as follows: For any user, a feature set consisting of all ad tags and the interest weights corresponding to each ad tag is established as the user profile of that user. The obtained valid interest weights are updated to the interest weights of the corresponding advertising tags in the user's user profile, thereby realizing the real-time construction of the user's user profile.

[0013] Furthermore, the method for obtaining the inflection point curvature is as follows: Obtain the discrete coordinate point sequence of the sliding trajectory during the current browsing advertisement dwell time period; based on the coordinate difference between adjacent discrete coordinate points in the discrete coordinate point sequence, obtain the horizontal first derivative, horizontal second derivative, vertical first derivative, and vertical second derivative corresponding to each discrete coordinate point respectively. For any discrete coordinate point, the product of the horizontal first derivative and the vertical second derivative corresponding to the discrete coordinate point is taken as the first eigenvalue; the product of the horizontal second derivative and the vertical first derivative corresponding to the discrete coordinate point is taken as the second eigenvalue; and the absolute value of the difference between the first eigenvalue and the second eigenvalue is taken as the first analytical value of the discrete coordinate point. The second analytical value of the discrete coordinate point is obtained by adding the square of the first derivative of the horizontal direction to the square of the first derivative of the vertical direction. The ratio of the first analysis value to the second analysis value is taken as the curvature of the discrete coordinate point; Discrete coordinate points with local maxima are selected from various discrete coordinate points as inflection points, and the curvature corresponding to the inflection points is used as the inflection point curvature in the sliding trajectory.

[0014] Secondly, another embodiment of the present invention provides a user profile construction system for the Internet advertising recommendation process. The system includes: a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above methods.

[0015] The present invention has the following beneficial effects: This invention first obtains the interaction hesitation index of the currently viewed advertisement based on the number of words in the advertisement, the magnitude and change of the scrolling rate during the time the user stays on the advertisement, and the curvature of the inflection point in the scrolling trajectory. This accurately reflects the user's reading depth and behavioral indecision regarding the current advertisement content, which helps to accurately distinguish between effective reading and ineffective idle behavior. To eliminate interference from the external physical environment on user behavior data, the invention further obtains the spatial stability coefficient of the currently viewed advertisement based on the similarity of the direction and amplitude changes of the terminal acceleration vector between adjacent moments during the time the user stays on the advertisement. This accurately reflects the stable state of the physical environment in which the user is viewing the advertisement, helping to avoid misjudging passive accidental touches or physical jitter as active user interaction. Finally, based on the interaction hesitation index and the spatial stability coefficient, the following data is obtained: The system extracts the immersive attention weight of the currently viewed ad to accurately reflect the user's true attention level and deep interest feedback, excluding physical environmental interference. Considering that users' true interests are dynamically evolving and that high-frequency exposure leads to diminishing marginal utility, in order to maintain the continuity of interest history while reasonably suppressing interest fatigue caused by repeated exposure, the effective interest weight of the currently viewed ad is updated based on the immersive attention weight, the historical interest weight corresponding to the ad tag of the currently viewed ad, and the cumulative exposure within a preset historical time period. This ensures that the final output weight includes both immediate real interactive feedback and takes into account historical accumulation and fatigue decay, thereby accurately constructing a user profile. This provides objective and high-precision data support for subsequent personalized ad recommendation engines, effectively avoiding information cocoons and improving the matching accuracy of recommendation results. Attached Figure Description

[0016] Figure 1 A schematic flowchart illustrating a user profile construction method in the process of internet advertising recommendation, provided as an embodiment of the present invention; Figure 2 This is a structural diagram of a user profile construction system in the process of internet advertising recommendation, provided as an embodiment of the present invention. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, details the specific solution of the user profile construction method and system provided by this invention in the process of internet advertising recommendation.

[0018] Example 1: This invention proposes a method for constructing user profiles in the process of internet advertising recommendation. Please refer to [link / reference]. Figure 1 The diagram illustrates a user profile construction method in the internet advertising recommendation process according to an embodiment of the present invention. The method includes the following steps: Step S1: Real-time collection of terminal acceleration vector, ad tags, number of characters, scrolling trajectory and scrolling speed when users browse ads.

[0019] Specifically, this embodiment uses a target user as an example for analysis, and all subsequent user references refer to this target user. To accurately depict the user's interaction state and environmental interference during ad browsing, and thus achieve precise discovery of the user's true interest in the current ad, this embodiment performs real-time multi-dimensional data collection on user browsing behavior and ad attributes. Specifically, it uses the inertial measurement unit (e.g., accelerometer) built into the user's terminal device to collect the terminal acceleration vector in real time while browsing ads, reflecting the terminal device's motion in three-dimensional space. Based on this, it assesses the degree of physical interference from the external environment on the user. Furthermore, by reading the pre-configured material metadata of the ad recommendation system, it obtains the ad tags of the ads currently being viewed by the user, representing the industry dimension or content category of the current ad (e.g., digital products, beauty and personal care). (Skin and sports / leisure, etc.) are the basic identifiers for subsequent accurate updates of the user profile's corresponding dimension weights; by directly reading the text attributes of the advertising material source file or calling the optical character recognition algorithm (public knowledge, not elaborated here), the number of characters in the advertisement currently being viewed by the user is obtained, objectively reflecting the information carrying capacity and reading complexity of the current advertisement; by listening to the touch event sequence of the user's terminal device touch screen, the sliding trajectory and sliding speed of the user when browsing the advertisement are collected in real time. Specifically, the sliding trajectory is generated by fitting the coordinate point sequence of the user's finger moving continuously on the screen, and the sliding speed is calculated by combining the time difference between adjacent touch coordinates, so as to depict the user's real behavioral feedback to the current advertisement.

[0020] It should be noted that, in order to ensure the accurate alignment of multimodal data in the time domain and avoid logical errors in subsequent temporal feature analysis (such as the correlation analysis between sudden action changes and sliding states) due to time offset, this embodiment sets the terminal acceleration vector, advertising label, text quantity, sliding trajectory and sliding rate to be collected synchronously. The underlying hardware data acquisition frequency is set to 60Hz to ensure that it can capture the slight shaking of the device and the user's short sliding operations delicately and continuously. Implementers can adjust the data acquisition frequency according to the actual computing power of the terminal device and the system power consumption requirements, which is not limited here.

[0021] Step S2: Based on the number of words in the ad currently being viewed by the user, the magnitude and change of the scrolling rate during the time the user stays on the ad, and the curvature of the inflection point in the scrolling trajectory, obtain the interaction hesitation index of the ad currently being viewed by the user.

[0022] Specifically, considering that traditional ad recommendations rely solely on ad dwell time, often failing to distinguish between "effective reading" and "ineffective browsing," this study aims to accurately depict users' true cognitive processing state and deeper interaction intentions while browsing ads. Based on the amount of text in the currently viewed ad, the magnitude and variation of the scrolling rate during the ad's dwell time, and the curvature of inflection points in the scrolling trajectory, an interaction hesitation index is obtained. This index accurately reflects the user's depth of reading and level of behavioral indecision regarding the current ad content, facilitating the precise allocation of subsequent interest weights. A higher interaction hesitation index indicates a deeper engagement with the ad content and a more complex decision-making process.

[0023] Preferably, in one feasible method of this embodiment, the method for obtaining the interaction hesitation index is as follows: First, obtain the mean and standard deviation of the scrolling speed during the current viewing time of the advertisement, and use them as the analysis mean and analysis standard deviation respectively; when the analysis mean and analysis standard deviation are smaller, it indicates that the user's finger scrolls slowly and the speed change is stable, and the user's visual focus is more likely to be in a focused reading state; at the same time, when the number of words in the advertisement currently viewed by the user is larger, it indicates that the information load of the current advertisement is larger, and the user needs to pay more attention; in order to quantify the user's actual reading and digestion ability of the current advertisement, the information acceptance level of the advertisement currently viewed by the user is obtained based on the number of words, analysis mean, and analysis standard deviation, which accurately reflects the user's ability to receive effective information of the current advertisement in the current scrolling state; the greater the information acceptance level, the more focused and efficient the user is in the information reception state; The specific method for obtaining information acceptability is as follows: the number of words, the mean of the analysis, and the standard deviation of the analysis are all standardized and mapped to a unified dimensionless interval (i.e., Within this framework, the differences in dimensions caused by different physical units among the parameters are eliminated, preventing direct calculations from leading to invalid physical meanings. The standardization process involves dividing the number of characters, the analytical mean, and the analytical standard deviation by their respective preset reference upper limits. These preset reference upper limits are set by analyzing the extreme values ​​of historical statistical data from advertising materials to ensure they cover the browsing scenarios of the vast majority of normal users, avoiding excessive compression of the mapping interval due to extreme outliers. It should be noted that if a parameter exceeds its corresponding preset reference upper limit, that parameter is truncated to its corresponding preset reference upper limit. The sum of the standardized analytical mean and standard deviation is used as the sliding stability; a lower sliding stability indicates a smoother sliding state. To intuitively establish the correlation between information load and sliding state, the ratio of the standardized number of characters to the sliding stability is used as the user's information acceptance level for the currently viewed advertisement. It should be noted that, in order to avoid a mathematical deadlock in the ratio calculation due to the denominator being zero, caused by the analysis mean and standard deviation being zero, the sum of the standardized analysis mean, the standardized analysis standard deviation, and a preset minimum positive number (e.g., 0.0001, without limitation) is used as the denominator (i.e., as the final sliding stability). In addition, if the user does not perform any sliding operations during the time period of the currently viewed advertisement, the mean of the sliding rate, the standard deviation of the sliding rate, and the inflection point curvature of the sliding trajectory are all directly set to 0 by default. Furthermore, considering that sharp bends in the swipe trajectory usually indicate visual review, repeated confirmation, or operational pauses by the user, in order to comprehensively evaluate the intensity of this behavioral feedback, the sum of the curvatures of all inflection points in the swipe trajectory during the current ad viewing time is obtained as the user's level of hesitation regarding the currently viewed ad. A higher level of hesitation indicates stronger user behavioral indecision and decision-making uncertainty. Specifically, the inflection point curvature is obtained as follows: A sequence of discrete coordinate points of the swipe trajectory during the current ad viewing time is obtained; based on the coordinate differences between adjacent discrete coordinate points in the sequence, the horizontal first derivative, horizontal second derivative, vertical first derivative, and vertical second derivative are obtained for each discrete coordinate point; for any discrete coordinate point, the horizontal first derivative is summed with... The product of the vertical second derivative is used as the first eigenvalue; the product of the horizontal second derivative and the vertical first derivative corresponding to the discrete coordinate point is used as the second eigenvalue; the absolute value of the difference between the first and second eigenvalues ​​is used as the first analytical value of the discrete coordinate point; the square of the horizontal first derivative and the square of the vertical first derivative corresponding to the discrete coordinate point is added together and the square of the third is taken to obtain the second analytical value of the discrete coordinate point; the ratio of the first analytical value to the second analytical value is used as the curvature of the discrete coordinate point; discrete coordinate points with local maxima of curvature are selected from all discrete coordinate points as inflection points, and the curvature corresponding to the inflection points is used as the inflection point curvature in the sliding trajectory; the above method for obtaining the inflection point curvature discretizes the continuous trajectory, enabling low-latency and high-precision extraction of the abrupt bending features of finger operations; Finally, to extract genuine interactive feedback while ensuring effective information reception and avoiding misjudging ineffective operations such as rapid swiping as effective interactions, the system obtains the product of information acceptance level and hesitation level. Considering that if a user operates with extremely slow movements and a variance infinitely close to zero, zero-padding to prevent errors would cause the product to spike dramatically, directly introducing this feature into the scale library would lead to severe distortion of the global comparison anchor point, causing the feature mapping result of normal browsing with deep intent to become excessively small and lose its discriminative power. Therefore, the system determines whether the product of information acceptance level and hesitation level is greater than a preset product upper limit (in this embodiment, the preset product upper limit is set to 100; implementers can adjust this according to their business scenarios). If it is greater, the product of information acceptance level and hesitation level is truncated to the preset product upper limit to obtain the target product; if it is not greater, the product of information acceptance level and hesitation level is directly used as the target product. After completing the security blocking and interception control of the estimated data, the system can continuously output intent information with normal discriminative power to the backend link. Furthermore, the result of performing max-min normalization on the target product is used as the interaction hesitation index for the ad currently being viewed by the user, thus smoothly mapping the value of the interaction hesitation index to... Within the interval. The dataset corresponding to the max-min normalization method here is a pre-established global benchmark dataset. This dataset is composed of the product of all information acceptance and hesitation levels generated by a large number of sample users in historical ad browsing scenarios, thus providing a stable and universally applicable mapping benchmark and avoiding distortion of the normalization scale due to the sparsity of historical behavioral data for a single user. Furthermore, to avoid the denominator in the max-min normalization method being 0 (i.e., the maximum and minimum values ​​are equal), a preset very small positive number (e.g., 0.0001, not limited here) is added when calculating the denominator (maximum value minus minimum value) to ensure the absolute safety of the formula calculation and the stability of the program operation. The max-min normalization method is well-known and will not be elaborated further.

[0024] Step S3: Based on the similarity of the direction and the change in amplitude of the terminal acceleration vector between adjacent moments within the time period of the advertisement currently viewed by the user, obtain the spatial stability coefficient of the advertisement currently viewed by the user.

[0025] Specifically, considering that users' daily lives (such as when on a bus, subway, or walking) are subject to physical vibrations that cause the terminal device to shake, leading to passive finger tremors or accidental touches, the trajectory bends caused by this physical vibration do not represent subjective hesitation in the user's interaction. In order to eliminate the interference of the external environment on user behavior data, the spatial stability coefficient of the currently viewed advertisement is obtained based on the similarity of the direction and amplitude changes of the terminal acceleration vector between adjacent moments within the time period of the currently viewed advertisement. This accurately reflects the stable state of the physical environment in which the user is viewing the current advertisement. The larger the spatial stability coefficient, the more stable the environment in which the user is located, and the more accurately the swiping interaction data can reflect the user's subjective intention.

[0026] Preferably, in one feasible implementation of this embodiment, the method for obtaining the spatial stability coefficient is as follows: To ensure the effectiveness of temporal feature calculation and avoid computational deadlock caused by a very small number of samples, the time within the currently viewed advertisement dwell time period is taken as the target time. When the number of target times is greater than or equal to a preset threshold, it indicates that sufficient terminal motion samples representing the environmental state have been accumulated. At this time, for any target time, the result of normalizing the cosine value of the terminal acceleration vector of the target time and its previous adjacent target time is taken as the directional stability value of the target time, accurately reflecting the consistency of the motion direction at the target time; the larger the directional stability value, the smaller the deflection of the terminal device's motion direction at the target time; wherein, in this embodiment, through The above cosine values ​​are normalized, where c is the cosine value; the range of values ​​for ensuring directional stability is [value missing]. To avoid logical deviations in subsequent calculations; in addition, when calculating the cosine value of the terminal acceleration vector, if the vector magnitude is zero, a preset minimum positive number (e.g., 0.0001) is added to each dimension component of the corresponding vector to prevent the denominator from being zero when calculating the cosine value; this embodiment sets the preset quantity threshold to 2 to ensure that an effective temporal state comparison can be formed. Implementers can set the preset quantity threshold according to the specific algorithm window requirements, which is not limited here; Considering that the average motion state over a short period of time represents the basic attitude of the terminal device (such as being stationary or moving at a constant speed), in order to quantify the deviation of a single point in time from the basic attitude, the average vector of the terminal acceleration vectors at all target moments is used as the overall analysis vector. Then, the magnitude of the difference vector between the terminal acceleration vector at the target moment and the overall analysis vector is subtracted to obtain the acceleration change analysis value at the target moment, which accurately reflects the degree of abrupt change in motion amplitude at the target moment. The smaller the acceleration change analysis value, the closer the motion amplitude of the terminal device at the target moment is to the overall basic attitude. To comprehensively evaluate the stability of motion on a macroscopic time scale, the ratio of the mean of the directional stability values ​​at all target moments to the mean of the acceleration change analysis values ​​is used as the trend stability coefficient. This accurately reflects the degree to which the overall direction of the terminal equipment is consistent and deviates little from the basic posture. The larger the trend stability coefficient, the more stable the macroscopic motion trend of the current terminal equipment. It should be noted that, in order to avoid the mean of the acceleration change analysis values ​​being 0 (i.e., the equipment is in absolute stillness or uniform linear motion), which would result in a zero denominator for the ratio calculation, this embodiment uses the sum of the mean of the acceleration change analysis values ​​and a preset minimum positive number (e.g., 0.0001, which is not limited) as the denominator for calculating the trend stability coefficient. In addition, considering that macroscopic averages often fail to reflect instantaneous high-frequency vibrations, in order to capture the intensity of vibrations at the micro level (i.e., between adjacent very short moments), the mean of the magnitude of the difference vector of the terminal acceleration vectors of any two adjacent target moments is used as the difference analysis value to accurately reflect the instantaneous displacement change rate of the equipment at adjacent moments; the smaller the difference analysis value, the more coherent and smooth the movement of the equipment on the micro time scale. To comprehensively integrate macro-trend stability and micro-instantaneous smoothness, and thus obtain the ratio of trend stability coefficient to difference analysis value; it should be noted that, to avoid the difference analysis value being 0, resulting in a zero denominator, this embodiment adds the difference analysis value to a preset minimum positive number (e.g., 0.0001, not limited) as the denominator; considering that in the extreme environment of absolute terminal rest, the aforementioned preset minimum positive number as the denominator would cause the ratio to explode by tens of thousands of times, if directly substituted into subsequent normalization calculations, it would become the global upper limit determining the maximum and minimum paradigms, severely suppressing the feature resolution band width between normal slight disturbances and obvious jitter; therefore, it is determined whether the ratio is greater than the preset ratio upper limit (in this embodiment, the preset ratio upper limit is set to 50, and the implementer can calibrate it in conjunction with the extreme value of normal business jitter); if it is greater, the ratio is truncated to the preset ratio upper limit, obtaining The target ratio is obtained; if it is not greater than the target ratio, it is directly used as the target ratio. After the capping interception is added, the occasional large deviation value is blocked before entering the pool for comparison, and the normal behavioral feature data can be distributed into the effective measurement boundary. Finally, the result of the target ratio is processed by the maximum and minimum normalization and used as the spatial stability coefficient of the advertisement currently viewed by the user. The dataset corresponding to the maximum and minimum normalization method here is a pre-established global benchmark dataset. This dataset is composed of the ratio results of all trend stability coefficients and difference analysis values ​​generated by a large number of sample users in historical advertisement viewing scenarios. This provides a stable and universal mapping benchmark and avoids the distortion of the normalization scale caused by the sparsity of historical behavioral data of a single user (similarly, when calculating the maximum value minus the minimum value of the denominator, the maximum and minimum normalization method here will also add a preset minimum positive number to prevent division by zero anomalies).

[0027] When the number of target moments is less than a preset threshold, it indicates that the current advertisement is in the very short instant of its initial playback (e.g., less than 0.03 seconds), making it impossible to extract effective temporal fluctuation features. In this case, a preset stability coefficient is used as the spatial stability coefficient of the advertisement currently viewed by the user to prevent the algorithm from failing to output results due to insufficient samples during the cold start phase. In this embodiment, the preset stability coefficient is set to 1, which assumes that during the extremely short initial browsing phase, the user is in an absolutely stable and ideal environment, and no additional environmental penalty is imposed on their interaction behavior due to missing data. Implementers can set the preset stability coefficient according to the degree of business conservatism, which is not limited here.

[0028] Step S4: Based on the interaction hesitation index and spatial stability coefficient, obtain the immersive attention weight of the advertisement currently being viewed by the user.

[0029] Specifically, a higher interaction hesitation index indicates a greater depth of user indecision and decision-making regarding the current advertisement, representing a higher subjective willingness to pay attention and interact. A higher spatial stability coefficient indicates a more stable physical environment for the user, reducing the likelihood of the captured hesitation behavior being affected by external disturbances. To refine the user's true interest while eliminating physical environmental interference, the product of the interaction hesitation index and the spatial stability coefficient is used as the immersive attention weight for the currently viewed advertisement. This accurately reflects the user's true attention level to the current advertisement in a stable environment, facilitating subsequent precise and dynamic updates to the interest profile of the corresponding ad tag. A higher immersive attention weight indicates a stronger and more genuine deep interest response from the user to the current advertisement, undisturbed by environmental factors. The interaction hesitation index and spatial stability coefficient both have a range of values ​​within the specified range. Therefore, the range of values ​​for the focus weight is constrained. Within the interval, not only is the consistency of physical dimensions guaranteed, but the redundancy of algorithm calculations is also greatly reduced.

[0030] Step S5: Based on the immersive attention weight, the historical interest weight corresponding to the ad tag of the ad currently viewed by the user, and the cumulative exposure within a preset historical time period, update the effective interest weight of the ad currently viewed by the user to build a user profile.

[0031] Specifically, considering that user interests are not static but are profoundly influenced by long-term historical preferences and can also lead to aesthetic fatigue due to frequent recent exposure, in order to dynamically and comprehensively depict the evolution of user interests, the effective interest weight of the ads currently viewed by the user is updated based on the weight of immersive attention, the historical interest weight corresponding to the ad tags of the ads currently viewed by the user, and the cumulative exposure within a preset historical time period. This ensures that the final output weight includes both instantaneous feedback and historical accumulation and fatigue suppression, thereby accurately constructing user profiles and providing high-precision data support for subsequent personalized ad recommendation engines, effectively improving the rationality of ad recommendations.

[0032] Preferably, in one feasible embodiment, the method for obtaining the effective interest weight is as follows: Considering that when the same type of advertisement is repeatedly exposed to users too many times, the user's novelty will decrease, i.e., the marginal utility will decrease, in order to quantify and specifically suppress the decline in new interest caused by frequent exposure, an exposure decay factor is obtained based on the cumulative exposure volume, and the repeated exposure feature is penalized and weighted down; wherein, the cumulative exposure volume and the exposure decay factor are negatively correlated; the formula for obtaining the exposure decay factor is: In the formula, exp is an exponential function with the natural constant as the base; S is the cumulative exposure. To preset the attenuation control constant, this embodiment sets it to 50, ensuring that the weight decreases gradually and reasonably with increasing exposure, avoiding a sharp drop in weight due to a very small number of exposures. Implementers can set this value based on their experience in judging user fatigue cycles in their business scenarios. The specific number of views is not limited here; it should be noted that the cumulative exposure is the total number of times a user views an ad with the same ad tag as the currently viewed ad within a preset historical time period; in this embodiment, the preset historical time period is set to 30 days to ensure that it can cover the user's recent effective interest memory period, while removing historical exposure data that is too old and invalid. The implementer can set the length of the preset historical time period according to the consumption of system storage resources, which is not limited here. Considering the historical continuity of users' genuine interests, fatigue decay should only apply to the current instantaneous feedback increment, rather than forcibly deducting the accumulated historical preference base. Therefore, the initial interest weight of the currently viewed advertisement is obtained by multiplying the immersive attention weight, the preset reference weight, and the exposure decay factor, and then adding this to the historical interest weight. This ensures that the local decay factor is only effective for this new interaction, normally inheriting the structured immersive weight base formed in the past. In this embodiment, the preset reference weight is set to 0.1 to ensure that the adjustment range of a single interaction on the overall profile is within a reasonable range, preventing drastic fluctuations in interest tags. Implementers can set the preset reference weight according to the needs of the business scenario regarding the speed of interest iteration; no limitation is imposed here. To prevent the user's interest weight from accumulating infinitely over time, leading to out-of-bounds values ​​and logical errors, the minimum value between the initial interest weight and the constant 1 is used as the effective interest weight of the currently viewed advertisement, strictly constraining the interest weight within the upper limit of 1. It should be noted that if a user views an ad with a corresponding ad tag for the first time, the historical interest weight corresponding to that ad tag will be initialized to 0. When updating the interest of this dimension in the next round, the effective interest weight calculated this time will be used as the historical interest weight for the next round of calculation.

[0033] Preferably, in one feasible way of this embodiment, the method for constructing a user profile is as follows: for any user, a feature set consisting of all advertising tags and the interest weights corresponding to each advertising tag is established as the user profile of that user, thus completing the structured and digital expression of the user's multidimensional interest preferences from the data structure level; in order to ensure that the recommendation system always performs accurate ad matching based on the user's latest cognitive state, the obtained effective interest weights are then updated to the interest weights of the corresponding advertising tags in the user's user profile (i.e., overwriting its original historical interest weights), thereby realizing the real-time construction of the user's user profile and forming a complete closed loop from real-time data collection, analysis and calculation to adaptive iteration of the profile.

[0034] In summary, this embodiment collects real-time data on the user's terminal acceleration vector, ad tags, text quantity, swipe trajectory, and swipe speed while browsing advertisements; it obtains an interaction hesitation index based on the text quantity, swipe speed magnitude and changes, and swipe trajectory inflection point curvature; it obtains a spatial stability coefficient based on the similarity of the direction and amplitude changes of the terminal acceleration vector at adjacent time points; it obtains an immersive attention weight based on the interaction hesitation index and spatial stability coefficient; and finally, it updates the effective interest weight based on the immersive attention weight, historical interest weight, and cumulative exposure to construct a user profile. This invention effectively eliminates physical environmental interference, accurately depicts the user's true interaction intentions, and suppresses repeated exposure fatigue, significantly improving the objectivity and accuracy of user profile construction.

[0035] Example 2: This invention also proposes a user profile construction system for the internet advertising recommendation process; please refer to [link / reference]. Figure 2 The diagram illustrates a user profile building system structure for an internet advertising recommendation process according to an embodiment of the present invention. The system includes: a data acquisition module 10, an interaction hesitation index acquisition module 20, a spatial stability coefficient acquisition module 30, an immersion attention weight acquisition module 40, and a user profile building module 50.

[0036] The data acquisition module 10 is used to collect in real time the terminal acceleration vector, ad labels, number of characters, scrolling trajectory and scrolling speed when the user browses the advertisement.

[0037] The interaction hesitation index acquisition module 20 is used to acquire the interaction hesitation index of the user's currently viewed advertisement based on the number of words in the advertisement currently viewed by the user, the magnitude and change of the scrolling rate during the time period of the currently viewed advertisement, and the curvature of the inflection point in the scrolling trajectory.

[0038] The spatial stability coefficient acquisition module 30 is used to acquire the spatial stability coefficient of the advertisement currently viewed by the user based on the similarity of the direction and the change of the amplitude of the terminal acceleration vector between adjacent moments within the time period of the advertisement currently viewed by the user.

[0039] The Immersive Attention Weight Acquisition Module 40 is used to obtain the immersive attention weight of the advertisement currently being viewed by the user based on the interaction hesitation index and the spatial stability coefficient.

[0040] The user profile building module 50 is used to update the effective interest weight of the ad currently viewed by the user in order to build a user profile based on the immersive attention weight, the historical interest weight corresponding to the ad tag of the ad currently viewed by the user, and the cumulative exposure within a preset historical time period.

[0041] It should be noted that the user profile construction system and the user profile construction method in the Internet advertising recommendation process provided in the above embodiments belong to the same concept. The specific implementation process can be found in the method embodiments, and will not be repeated here.

Claims

1. A method for constructing user profiles in the process of internet advertising recommendation, characterized in that, The method includes the following steps: Real-time data collection of terminal acceleration vector, ad tags, text quantity, scrolling trajectory, and scrolling speed when users browse ads; Based on the number of words in the ad currently being viewed by the user, the magnitude and change of the scrolling rate during the time the user stays on the ad, and the curvature of the inflection point in the scrolling trajectory, the interaction hesitation index of the ad currently being viewed by the user is obtained. Based on the similarity of the direction and the change in the amplitude of the terminal acceleration vector between adjacent moments within the time period during which the user is currently viewing the advertisement, the spatial stability coefficient of the advertisement currently being viewed by the user is obtained. Based on the interaction hesitation index and spatial stability coefficient, obtain the immersive attention weight of the advertisement currently viewed by the user; Based on the immersion attention weight, the historical interest weight corresponding to the ad tag of the ad currently viewed by the user, and the cumulative exposure within a preset historical time period, the effective interest weight of the ad currently viewed by the user is updated to build a user profile.

2. The user profile construction method in the internet advertising recommendation process as described in claim 1, characterized in that, The method for obtaining the interaction hesitation index is as follows: The mean and standard deviation of the scrolling rate during the current viewing time of the advertisement are obtained and used as the analysis mean and standard deviation, respectively. Based on the number of words in the advertisement currently being viewed by the user, the mean of the analysis, and the standard deviation of the analysis, the degree of information acceptance of the advertisement currently being viewed by the user is obtained; The sum of the curvatures of all inflection points in the scrolling trajectory during the current viewing time of the advertisement is used as the degree of hesitation of the user regarding the current advertisement. The result of multiplying the degree of information acceptance and the degree of hesitation, and then normalizing the result, is used as the interaction hesitation index of the advertisement currently being viewed by the user.

3. The user profile construction method in the internet advertising recommendation process as described in claim 2, characterized in that, The method for obtaining the degree of information acceptance is as follows: The number of characters, the mean of the analysis, and the standard deviation of the analysis are all standardized and mapped to a unified dimensionless interval. The normalized mean and standard deviation of the analysis are added together to determine the sliding stability. The ratio of the standardized number of characters to the degree of scrolling stability is used as the user's acceptance level of the information in the advertisement they are currently viewing.

4. The user profile construction method in the internet advertising recommendation process as described in claim 1, characterized in that, The method for obtaining the spatial stability coefficient is as follows: Use the time period during which the currently viewed advertisement is viewed as the target time; When the number of target moments is greater than or equal to the preset number threshold, for any target moment, the result of normalizing the cosine value of the terminal acceleration vector of the target moment and the previous adjacent target moment is used as the directional stability value of the target moment. The average vector of the terminal acceleration vectors at all target times is used as the overall analysis vector; The magnitude of the difference vector between the terminal acceleration vector at the target moment and the overall analysis vector is subtracted to obtain the acceleration change analysis value at the target moment. The ratio of the mean of the directional stability values ​​at all target times to the mean of the acceleration change analysis values ​​is used as the trend stability coefficient. The mean of the magnitudes of the difference vectors between the terminal acceleration vectors of any two adjacent target moments is used as the difference analysis value. The result of normalizing the ratio of the trend stability coefficient to the difference analysis value is used as the spatial stability coefficient of the advertisement currently viewed by the user. When the number of target moments is less than the preset threshold, the preset stability coefficient is used as the spatial stability coefficient of the advertisement currently viewed by the user.

5. The user profile construction method in the Internet advertising recommendation process as described in claim 1, characterized in that, The method for obtaining the immersion attention weight is as follows: The product of the interaction hesitation index and the spatial stability coefficient is used as the weight of the user's immersive attention to the advertisement currently being viewed.

6. The user profile construction method in the Internet advertising recommendation process as described in claim 1, characterized in that, The method for obtaining the cumulative exposure is as follows: The total number of times a user views an ad with the same ad tag as the currently viewed ad within a preset historical time period is used as the cumulative exposure.

7. The user profile construction method in the Internet advertising recommendation process as described in claim 1, characterized in that, The method for obtaining the effective interest weight is as follows: Based on the cumulative exposure, an exposure attenuation factor is obtained; the cumulative exposure and the exposure attenuation factor are negatively correlated. The initial interest weight of the ad currently viewed by the user is obtained by multiplying the immersive attention weight, the preset reference weight, and the exposure decay factor, and then adding the historical interest weight. The minimum value between the initial interest weight and constant 1 is taken as the effective interest weight of the advertisement currently viewed by the user.

8. The user profile construction method in the Internet advertising recommendation process as described in claim 1, characterized in that, The method for constructing user profiles is as follows: For any user, a feature set consisting of all ad tags and the interest weights corresponding to each ad tag is established as the user profile of that user. The obtained valid interest weights are updated to the interest weights of the corresponding advertising tags in the user's user profile, thereby realizing the real-time construction of the user's user profile.

9. The user profile construction method in the Internet advertising recommendation process as described in claim 1, characterized in that, The method for obtaining the inflection point curvature is as follows: Obtain the discrete coordinate point sequence of the sliding trajectory during the current browsing advertisement dwell time period; based on the coordinate difference between adjacent discrete coordinate points in the discrete coordinate point sequence, obtain the horizontal first derivative, horizontal second derivative, vertical first derivative, and vertical second derivative corresponding to each discrete coordinate point respectively. For any discrete coordinate point, the product of the horizontal first derivative and the vertical second derivative corresponding to that discrete coordinate point is taken as the first eigenvalue. The product of the horizontal second derivative and the vertical first derivative corresponding to the discrete coordinate point is taken as the second eigenvalue. The absolute value of the difference between the first feature value and the second feature value is taken as the first analysis value of the discrete coordinate point; The second analytical value of the discrete coordinate point is obtained by adding the square of the first derivative of the horizontal direction to the square of the first derivative of the vertical direction. The ratio of the first analysis value to the second analysis value is taken as the curvature of the discrete coordinate point; Discrete coordinate points with local maxima are selected from various discrete coordinate points as inflection points, and the curvature corresponding to the inflection points is used as the inflection point curvature in the sliding trajectory.

10. A user profile building system for internet advertising recommendation process, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the user profile construction method in the Internet advertising recommendation process as described in any one of claims 1-9.

Images