Precise advertisement putting method based on multi-dimensional user portraits

Abstract

The invention provides a precise advertisement putting method based on a multi-dimensional user portrait, and the method comprises the steps: analyzing the distribution of an overlook angle, a look-up angle and a head-up angle in the historical photographing data of a user, and obtaining a dynamic change vector of the photographing angle preference of the user through calculating the change rate of the use frequency of each angle in a time sequence; according to the method, dependency modes of users on different composition auxiliary tools are analyzed through a personalized grid pattern preference scoring matrix, the users are divided into a high dependency type group, a moderate dependency type group and a low dependency type group by adopting a clustering algorithm, and the activation frequency and the stay time of each group on a golden section point prompting function in a shooting process are counted; and forming a user composition auxiliary tool use behavior file.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for precise advertising based on multi-dimensional user profiles. Background Technology

[0002] The widespread adoption of mobile photography technology has made composition assist tools a key technology area for enhancing the user shooting experience. This area directly impacts the visual creative quality of hundreds of millions of users and the accuracy of advertising. As users' demands for photographic quality continue to rise, composition assist technology has become an indispensable core component of the mobile application ecosystem. Current composition assist advertising solutions mainly rely on static user tags and simple usage frequency statistics, lacking a deep understanding of user shooting behavior. These solutions cannot identify users' personalized needs in different shooting scenarios, resulting in a significant disconnect between ad recommendations and users' actual composition habits, leading to resource waste and a decline in user experience. The dynamic changes in users' shooting angle preferences are difficult to accurately capture and quantify. As users gradually try diverse angles such as overhead and low-angle shots, their demand for composition assist tools changes significantly. This shift in angle preference directly affects users' preference for composition grid line styles. The segmentation of composition grid line style preferences further leads to a differentiation in users' reliance on golden ratio point prompts. Some users may reduce their reliance on assist prompts after mastering multi-angle shooting techniques, while others may increase their need for precise positioning tools. The individual variations in reliance on the golden ratio highlight mean that traditional uniform advertising recommendation strategies cannot meet users' personalized needs, necessitating the establishment of a more refined dynamic matching mechanism. How to establish a dynamic optimization mechanism for composition assist tool advertising recommendations based on changes in shooting angle preferences, mapping angle selection habits to composition tool needs, and effectively improving the conversion rate of composition ads, has become a key issue in the personalized application of composition assist technology. Summary of the Invention

[0003] This invention provides a method for precise ad delivery based on multi-dimensional user profiles, mainly including:

[0004] By analyzing the distribution of top-down, bottom-up, and eye-level angles in users' historical shooting data, and calculating the rate of change of the frequency of use of each angle over time, a dynamic change vector of users' shooting angle preferences is obtained.

[0005] Based on the dynamic change vector of user's shooting angle preference, an angle-grid mapping relationship model is established. When the growth rate of the user's use of the overhead angle exceeds the preset threshold, the recommendation weight of the three-part grid pattern is activated. When the user's use of the eye-level angle increases, the matching priority of the golden spiral grid pattern is adjusted to obtain a personalized grid pattern preference score matrix.

[0006] By analyzing users' dependence patterns on different composition aids through a personalized grid style preference rating matrix, and using a clustering algorithm to divide users into three groups: high dependence, moderate dependence, and low dependence, the activation frequency and dwell time of the golden ratio prompt function for each group during the shooting process were statistically analyzed to form a user composition aid usage behavior profile.

[0007] Based on the user composition aid tool usage behavior profile, the user's reliance level on the golden ratio point prompt was determined. High-reliance users frequently use various grid lines and rely on the golden ratio point prompt to complete the composition. Moderate-reliance users occasionally use grid styles but can independently judge the composition effect. Low-reliance users rarely use aid tools and mainly rely on personal experience to complete the shooting.

[0008] After obtaining the user's golden ratio point prompt dependence level, the three dimensions of features—photo angle preference, grid style selection, and prompt tool usage time—are integrated. A comprehensive preference index is calculated through a feature weighted fusion algorithm to obtain the user's composition behavior feature vector. This vector reflects the behavioral differences of users with different dependence levels in each dimension. At the same time, the frequency of user switching between three angles—top, bottom, and eye level—is calculated as an angle diversity indicator.

[0009] Based on the user's composition behavior feature vector, a dynamic ad matching strategy is designed. By analyzing the matching features of angle diversity index and cue dependence in the user's comprehensive preference index, the ad recommendation category is determined. When the user displays high angle diversity and low cue dependence, professional composition tool ads are recommended. When the user displays a single angle but high cue dependence, basic auxiliary function ads are recommended, thus achieving accurate ad content selection and delivery strategy formulation.

[0010] Monitor ad click-through rates and conversion behavior, dynamically adjust the weights of three dimensions in the user composition behavior feature vector: shooting angle preference, grid style selection, and duration of use of prompt tools. At the same time, combine the real-time angle detection results of the user's current shooting scene to optimize the ad display strategy. When it is detected that the user is shooting from above and historical data shows that this angle is a new attempt, increase the display weight of composition tutorial ads to determine the final personalized ad recommendation results.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] This invention discloses a method for precise ad delivery based on multi-dimensional user profiles. By analyzing the angle distribution characteristics in users' historical shooting data, an angle-grid mapping model is established to obtain a personalized grid style preference rating matrix. Based on users' dependence patterns on composition aids, users are categorized and composition aid usage behavior profiles are created, thereby determining the user's dependence level on the golden ratio prompt. This invention integrates multi-dimensional features such as shooting angle preferences, grid style selection, and prompt tool usage duration to calculate user composition behavior feature vectors, and designs dynamic ad matching strategies based on these vectors. Through a real-time feedback mechanism and a Bayesian update algorithm, the feature vector weights are dynamically adjusted, and the ad display strategy is optimized in real-time based on the current shooting scene to achieve accurate personalized ad recommendations. This invention effectively improves ad matching accuracy and user experience, providing a new technical solution for precision marketing in photography applications. Attached Figure Description

[0013] Figure 1 This is a flowchart of a precise advertising delivery method based on multi-dimensional user profiles according to the present invention.

[0014] Figure 2 This is a schematic diagram of a precise advertising delivery method based on multi-dimensional user profiles according to the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] like Figure 1 -2, This embodiment of a precise advertising delivery method based on multi-dimensional user profiles may specifically include:

[0017] Step S101: Analyze the distribution of top-down angle, bottom-up angle, and eye-level angle in the user's historical shooting data. By calculating the rate of change of the frequency of use of each angle in the time series, obtain the dynamic change vector of the user's shooting angle preference.

[0018] The shooting timestamp and tilt angle value recorded by the camera gyroscope sensor are extracted from the user's historical shooting data for each photo. A tilt angle greater than 45 degrees with the camera pointing downwards is determined as a top-down angle; a tilt angle greater than 45 degrees with the camera pointing upwards is determined as a bottom-up angle; and a tilt angle between ±45 degrees is determined as a level angle. The shooting frequency of each angle type in different time periods is statistically analyzed to form an angle distribution feature matrix. For this angle distribution feature matrix, the time series is divided into continuous time windows. The rate of change of the frequency of each angle type within adjacent time windows is calculated. The rate of change is equal to the frequency of the current window minus the frequency of the previous window, then divided by the frequency of the previous window. If the rate of change exceeds a preset growth threshold, the weight coefficient of that angle increases by 0.1; if the rate of change is lower than a preset decrease threshold, the weight coefficient decreases by 0.1. The trend coefficient is obtained by accumulating the weight coefficient changes of each time window. Based on the trend coefficient and the weight coefficient of each angle type, the weight coefficient of each angle type is normalized so that the sum is 1, and a dynamic change vector of user shooting angle preference is constructed. Each element in this vector corresponds to an angle type, and the element value is equal to the product of the normalized weight coefficient and the trend coefficient of that angle, thus obtaining a dynamic change vector of user shooting angle preference that includes angle type weight and trend coefficient.

[0019] Specifically, camera gyroscope sensors are standard in modern smartphones and digital cameras, and they can detect the tilt angle of the device in three-dimensional space in real time.

[0020] In one possible implementation, when a user raises their phone to take a picture, the gyroscope continuously outputs the tilt angle of the device relative to the horizontal plane. This angle is typically measured in degrees and ranges from -90 to 90 degrees. When the angle exceeds 45 degrees and the camera lens is pointing towards the ground, the system determines it as a top-down shot; when the angle exceeds 45 degrees and the lens is pointing towards the sky, it is determined as a bottom-up shot; and between -45 degrees and 45 degrees, it is a level shot.

[0021] Specifically, the construction of the angle distribution feature matrix involves statistical analysis of a large amount of historical data. Suppose a user took 300 photos in the past month. The system divides this month into 30 time periods, each corresponding to one day. On the first day, the user took 10 photos: 6 from a top-down angle, 3 from a level angle, and 1 from a bottom-up angle. Therefore, the angle distribution for that day is 60% from top-down, 30% from level, and 10% from bottom. By statistically analyzing the 30 days of data, a 30-row, 3-column matrix is ​​formed, with each row representing one day and each column corresponding to the frequency of the three shooting angles.

[0022] In one embodiment, the time window is divided using a sliding window mechanism, with each window containing data for seven consecutive days. The calculation of the rate of change reflects the dynamic changes in users' shooting habits.

[0023] For example, if a user takes 20 shots from above in the first week, and this increases to 35 in the second week, the rate of change is calculated as ...35 - 20... divided by 20, which equals 0.75, representing a 75% increase. If the preset growth threshold is 0.5, the weighting coefficient for the overhead angle increases by 0.1. This cumulative mechanism allows the system to capture long-term trends in user shooting preferences.

[0024] It's important to note that normalization ensures the comparability of weights for different angle types. Assuming that after accumulating over multiple time windows, the original weight coefficients for overhead, eye-level, and upward-looking views are 0.8, 0.5, and 0.2 respectively, after normalization they become 0.53, 0.33, and 0.14. The trend change coefficient reflects the rate of change in user preferences; if a certain angle type has been frequently used and continues to increase recently, its trend change coefficient will be larger. The final dynamic change vector is obtained by multiplying the normalized weights by the trend coefficient. This design allows the system to consider both current usage preferences and the changing trends of those preferences, providing reliable data support for subsequent photo suggestions and optimizations.

[0025] Step S102: Establish an angle-grid mapping relationship model based on the dynamic change vector of user's shooting angle preference. When the growth rate of the user's overhead angle usage exceeds a preset threshold, activate the recommendation weight of the three-part grid pattern. When the user's eye-level angle usage frequency increases, adjust the matching priority of the golden spiral grid pattern to obtain a personalized grid pattern preference rating matrix.

[0026] Based on the angle type weights and trend coefficients in the dynamic vector of user shooting angle preferences, an angle-grid mapping relationship is established. The top-down angle corresponds to a rule-of-thirds grid pattern, where the image is divided into three equal parts horizontally and vertically to form a nine-square grid. The eye-level angle corresponds to a golden spiral grid pattern. The strength of the mapping relationship is equal to the trend coefficient of the corresponding angle multiplied by a normalized weight coefficient. For the strength of this mapping relationship, the usage frequency growth rate of the top-down angle is calculated. If this growth rate exceeds a preset threshold of 0.5, the recommended weight of the rule-of-thirds grid pattern is activated and set as the strength value of the top-down angle. The usage frequency percentage of the eye-level angle is calculated. If the percentage increases compared to the previous time window, the matching priority of the golden spiral grid pattern is adjusted to the strength value of the eye-level angle multiplied by the percentage increase. Based on the activated rule-of-thirds grid pattern recommended weight and the adjusted golden spiral grid pattern matching priority, a personalized grid pattern preference rating matrix is ​​constructed. Each row in the matrix represents a grid pattern, and each column represents the preference rating for that grid pattern. The rating of the rule-of-thirds grid is its recommended weight, and the rating of the golden spiral grid is its matching priority, resulting in a personalized grid pattern preference rating matrix.

[0027] Specifically, the rule of thirds grid, as the most basic and practical auxiliary tool in photographic composition, works by dividing the image into three equal parts in both the horizontal and vertical directions, forming two horizontal lines and two vertical lines. These four lines intersect to create four intersection points.

[0028] In one possible implementation, when a user shoots from a top-down angle, the subject often appears flat. In this case, the rule of thirds grid can help the user arrange the positions of elements in the frame. For example, when shooting a top-down photo of food, placing the plate at the intersection of the grid points can achieve a composition with a stronger sense of visual balance.

[0029] Specifically, the strength of the mapping relationship reflects the degree of correlation between a user's shooting habits and the grid style. Assuming a user's top-down angle trend coefficient is 0.8 and the normalized weighting coefficient is 0.53, the strength of the mapping relationship between the top-down angle and the rule of thirds grid is calculated as 0.8 multiplied by 0.53, which equals 0.424. This value indicates the degree to which the system recommends the rule of thirds grid when the user is using a top-down angle. The golden spiral grid style, based on the spiral curve formed by the Fibonacci sequence, is particularly suitable for portrait or landscape photography at eye level, because at eye level, the elements in the image usually have a clear hierarchy and visual guidance path.

[0030] It should be noted that the calculation of the frequency growth rate involves comparative analysis over time. When the system detects that a user's number of overhead shots has increased from 20 to 35 in the past week, reaching a growth rate of 75% and exceeding the preset threshold of 0.5, it triggers the activation mechanism for the recommendation weight of the three-part grid pattern. This activation is not a simple on / off switch, but rather sets the recommendation weight of this grid pattern to an intensity value of 0.424 based on the overhead angle, so that the system will prioritize this grid pattern in subsequent recommendations.

[0031] In one embodiment, changes in the frequency of use of the eye-level angle directly affect the matching priority adjustment of the golden spiral grid. Assuming the eye-level angle usage rate was 30% in the previous time window and increases to 45% in the current window (an increase of 0.5), the system multiplies the eye-level angle intensity value (0.33) by the increase in rate (0.5) to obtain a new matching priority of 0.165 for the golden spiral grid. This dynamic adjustment mechanism allows grid recommendations to respond in real-time to changes in user shooting habits. The construction process of the personalized grid style preference scoring matrix integrates the recommendation weights and matching priorities of various grid styles into a unified scoring system. The matrix may be presented in a 2x1 format, with the first row corresponding to the ternary grid and a score of 0.424 (the activated recommendation weight); the second row corresponds to the golden spiral grid and a score of 0.165 (the adjusted matching priority). This quantitative scoring mechanism provides a clear decision-making basis for subsequent intelligent grid recommendations, enhancing the personalization of the user shooting experience.

[0032] Step S103: Analyze users' dependence patterns on different composition aids through a personalized grid style preference rating matrix. Use a clustering algorithm to divide users into three groups: high dependence, moderate dependence, and low dependence. Statistically analyze the activation frequency and dwell time of the golden ratio point prompt function for each group during the shooting process to form a user composition aid usage behavior profile.

[0033] By using a personalized grid style preference rating matrix, the total score of a user's overall dependence on the composition aid tool is obtained by adding the scores of the ternary grid and the golden spiral grid in the matrix. If the total score is greater than a preset upper threshold of 0.8, the user is marked as highly dependent; if it is less than a preset lower threshold of 0.3, the user is marked as low dependent; and if it is between the two thresholds, the user is marked as moderately dependent. Based on the user dependence type labels, user rating matrix data with the same labels are extracted. The K-means clustering algorithm is used to further subdivide users within each dependence type. The number of clusters is set to 3. The two-dimensional feature vectors constructed by the user's ternary grid score and the golden spiral grid score are used for clustering calculation to obtain the subdivision results of the three user groups: highly dependent, moderately dependent, and low dependent. Based on the segmentation results of the three user groups, the activation frequency is obtained by dividing the number of times users in each group activate the golden section point marker in the grid auxiliary lines on the shooting interface by the total number of shots. The duration of the user's time on the golden section point display interface after each activation is recorded, and the arithmetic mean of all activation durations is calculated to obtain the average dwell time. The group type, activation frequency and average dwell time data are integrated to form a user composition assistance tool usage behavior profile.

[0034] Specifically, the personalized grid style preference rating matrix, as the core data structure for quantifying users' composition habits, essentially transforms users' preferences for different grid styles into a calculable numerical representation.

[0035] In one possible implementation, when the user's score for the rule of thirds grid is 0.424 and the score for the golden spiral grid is 0.165, the sum of the two is 0.589. This total score directly reflects the user's overall reliance on the composition aid tool, falling within the threshold range of 0.3 to 0.8, and is therefore marked as moderately dependent.

[0036] Specifically, the dependency type is categorized based on the frequency and degree of user reliance on grid lines during actual shooting. High-dependency users typically enable grid lines in every shot, achieving a total score above 0.8; low-dependency users rarely use auxiliary tools, relying more on intuitive composition, resulting in a total score below 0.3; moderately dependent users selectively use grid lines in specific scenarios. This classification method makes subsequent personalized recommendations more accurate. The application of the K-means clustering algorithm in this scenario requires constructing appropriate feature vectors. Each user's feature vector consists of two dimensions: a ternary grid score and a golden spiral grid score.

[0037] For example, if a user's feature vector is [0.424, 0.165], the algorithm calculates the Euclidean distance between different user feature vectors to group similar users into the same category. The clustering process iteratively optimizes and continuously adjusts the cluster center positions until the user features within each category are sufficiently similar and the differences between categories are sufficiently significant.

[0038] In one embodiment, the golden ratio point is an important concept in composition theory, located at 0.618 times the width and height of the image. When a user activates the golden ratio point marker function, the shooting interface will display prominent markers at these key locations to help the user place the main subject element at the visual focal point. User behavior data for activating this function is recorded; for example, a highly dependent user activated the golden ratio point marker 78 times out of 100 shots, with an activation frequency of 0.78.

[0039] It's important to note that the dwell time statistics reflect the actual depth of user engagement with the aids. For example, if a user activates the golden ratio marker and spends an average of 8.5 seconds on the screen to adjust their composition, while another user only spends 2.3 seconds, this difference indicates that the former relies more heavily on the aids for precise composition. By integrating data such as the activation frequency of 0.78 and the average dwell time of 8.5 seconds, combined with the user group type tag "high dependence," a complete user profile of composition aid usage behavior is created. This profile not only records user habits but also provides data support for subsequent personalized feature optimization.

[0040] Step S104: Determine the user's reliance level on the golden ratio point prompt based on the user's composition aid tool usage behavior profile. High-reliance users frequently use various grid lines and rely on the golden ratio point prompt to complete the composition. Moderate-reliance users occasionally use grid styles but can independently judge the composition effect. Low-reliance users rarely use aid tools and mainly rely on personal experience to complete the shooting.

[0041] Based on user behavior profiles of composition aids, including group type, activation frequency, and average dwell time, usage characteristics of the golden ratio point prompt function are extracted. If the activation frequency exceeds a preset high-frequency threshold and the average dwell time exceeds a preset upper limit, the user's reliance on the golden ratio point prompt is marked as high-level. If the activation frequency is between the preset high-frequency and low-frequency thresholds, it is marked as medium-level. If the activation frequency is below the preset low-frequency threshold, it is marked as low-level. Based on these golden ratio point prompt reliance levels and combined with grid pattern usage records in the user behavior profiles of composition aids, the ratio of the cumulative usage of various grid lines to the total number of shots is calculated to obtain the overall usage rate of the grid aid tool. The overall usage rate of high-level users exceeds a preset upper limit, the overall usage rate of medium-level users is between the preset upper and lower limits, and the overall usage rate of low-level users is below the preset lower limit. Based on the overall usage rate of the grid assistance tool and the level of dependence on the golden section point prompt, the user's golden section point prompt dependence level is determined. Users with a high level and an overall usage rate exceeding the preset upper limit are determined to be highly dependent; users with a medium level and an overall usage rate in the middle range are determined to be moderately dependent; and users with a low level and an overall usage rate below the preset lower limit are determined to be low dependent.

[0042] Specifically, the user composition assistance tool uses behavioral profiles as the core data carrier to record users' shooting habits, which includes multi-dimensional information such as group type, activation frequency, and average dwell time.

[0043] In one possible implementation, the activation frequency reflects the proportion of times a user activates the golden ratio point tooltip within a certain period. For example, if a photography enthusiast took 200 photos in the past month and activated the golden ratio point tooltip 150 times, the activation frequency would be 0.75. This value far exceeds the preset high-frequency threshold of 0.6, indicating that the user has a strong reliance on the tooltip.

[0044] Specifically, the average dwell time provides a deeper reflection of the user's actual usage of the assistive tool. When a user activates the golden ratio prompt, the system starts timing until the user completes the composition and presses the shutter or deactivates the prompt. A professional photographer, despite an activation frequency of 0.8, only had an average dwell time of 3 seconds, indicating that they only used the tool as a quick reference. In contrast, a photography beginner, with an activation frequency of 0.7, had an average dwell time of 15 seconds, reflecting that they needed more time to rely on the tool to adjust their composition.

[0045] It should be noted that the calculation of the overall utilization rate of grid auxiliary tools covers all types of compositional guidelines, including not only the golden ratio but also the rule of thirds grid, the golden spiral grid, and other auxiliary tools. This comprehensive evaluation method avoids the one-sidedness of a single indicator.

[0046] For example, a user may rarely use the golden ratio point as a cue, but frequently uses the rule of thirds grid, and their overall usage rate is still high. This indicates that the user has a strong overall reliance on composition aids, but prefers different types of guide lines.

[0047] In one embodiment, the final determination of dependency level employs a dual-judgment mechanism. The first judgment is based on the usage characteristics of the golden ratio cues, while the second judgment combines the overall usage of all auxiliary tools. This design takes into account the diversity of user habits. A seasoned photographer might frequently use grid lines to ensure vertical lines when shooting architecture, achieving a combined usage rate of 0.85, but completely ignore the golden ratio cues when shooting portraits, with an activation frequency of only 0.1. Through this dual judgment, the system categorizes this user as moderately dependent, acknowledging their need for auxiliary tools in specific scenarios while also recognizing their ability to compose independently. Highly dependent users are characterized by comprehensive reliance on all types of auxiliary tools, exhibiting not only high activation frequency and long dwell time, but also a significant decrease in composition quality after disabling auxiliary tools. Moderately dependent users demonstrate selective use, flexibly deciding whether to enable auxiliary tools based on the shooting scenario. Lowly dependent users primarily rely on personal experience and aesthetic intuition for composition, using auxiliary tools only occasionally as a reference. This classification provides a precise user profile foundation for subsequent personalized feature recommendations and user experience optimization.

[0048] Step S105: After obtaining the user's golden ratio point prompt dependence level, the three dimensions of features—photo angle preference, grid style selection, and prompt tool usage time—are integrated. A comprehensive preference index is calculated using a feature weighted fusion algorithm to obtain the user's composition behavior feature vector. This vector reflects the behavioral differences of users with different dependence levels in each dimension. At the same time, the frequency of switching between the user's top-down, bottom-up, and eye-level angles is calculated as an angle diversity indicator.

[0049] After obtaining the user's golden ratio point suggestion dependency level, the angle weight value in the dynamic change vector of shooting angle preference, the grid score value in the personalized grid style preference rating matrix, and the average dwell time data in the composition assistance tool usage behavior profile are extracted. Each data point is normalized by dividing it by the maximum value of that dimension, resulting in standardized feature values ​​ranging from 0 to 1. Based on the standardized feature values, a weighted summation method is used to calculate the comprehensive preference index. The weights of the angle preference dimension, grid selection dimension, and usage time dimension are set to preset first, second, and third weight values, respectively, and the sum of the three weight values ​​is 1. The comprehensive preference index is obtained by multiplying the standardized feature values ​​of each dimension by their corresponding weights and then summing them. Based on the comprehensive preference index, and combining the standardized feature values ​​of the three dimensions and the user dependence level, the angle type change of the user in two adjacent photos in a continuous shooting sequence is statistically analyzed. If the previous photo is a top view and the next photo is a bottom view or a level view, it is counted as one switch. The cumulative number of switches is divided by the shooting sequence length minus 1 to obtain the angle switch rate. A user composition behavior feature vector containing the comprehensive preference index, the three standardized feature values ​​and the angle switch rate is constructed to form a multi-dimensional user profile that reflects the behavioral differences of users with different dependence levels in various dimensions.

[0050] Specifically, normalization is a key step in data preprocessing, and its purpose is to transform characteristic values ​​with different dimensions and numerical ranges into a unified standard interval.

[0051] In one possible implementation, the angle weights might originally vary between 0 and 1, while the average dwell time could range from a few seconds to tens of seconds, and the grid score also has its specific range. By dividing by the maximum value of each dimension, all features are mapped to the interval between 0 and 1, making the subsequent weighted calculations more reasonable.

[0052] Specifically, suppose the original data of a highly dependent user is as follows: top-down angle weight value 0.8, three-part grid score 0.6, and average dwell time 20 seconds. If the maximum values ​​of this dimension are 1, 1, and 30 seconds respectively, after normalization, we get 0.8, 0.6, and 0.67. This processing eliminates the difference in units between different features, preventing a feature with a larger value from dominating the comprehensive calculation. The application of the weighted summation method reflects the difference in importance of different dimensions to the user's composition behavior. Angle preference reflects the user's shooting perspective habits, grid selection reflects the degree of preference for auxiliary tools, and usage time characterizes the depth of dependence on auxiliary functions. The setting of the three preset weight values ​​follows the constraint that the sum is 1, ensuring that the comprehensive preference index remains within the standard range of 0 to 1.

[0053] For example, when the first weight is 0.5, the second weight is 0.3, and the third weight is 0.2, the overall preference index of the above users is calculated as 0.8×0.5+0.6×0.3+0.67×0.2=0.714.

[0054] In one embodiment, the angle switching rate reveals the diversity of a user's shooting style. A continuous shooting sequence refers to a sequence of photos generated by a user in a single shooting activity, where the time interval between adjacent photos typically does not exceed a preset threshold. If a user takes 10 photos consecutively, in the order of top-down, top-down, eye-level, low-down, low-down, eye-level, top-down, top-down, eye-level, low-down, then the angle switching occurs between the 2nd and 3rd photos, the 3rd and 4th photos, the 5th and 6th photos, the 6th and 7th photos, and the 8th and 9th photos, for a total of 5 switching times. The angle switching rate is calculated as 5 divided by 9, approximately 0.56.

[0055] It's important to note that the construction of the user composition behavior feature vector integrates information from multiple dimensions. This vector might be represented as [0.714, 0.8, 0.6, 0.67, 0.56], corresponding to the comprehensive preference index, angle preference standardized value, grid selection standardized value, usage time standardized value, and angle switching rate, respectively. This vectorized representation allows for comparison and analysis of different users' composition behaviors in a multi-dimensional space. The establishment of multi-dimensional user profiles lays the foundation for personalized recommendations. Highly dependent users typically exhibit a high comprehensive preference index and a low angle switching rate, indicating a heavy reliance on auxiliary tools and a relatively fixed shooting style. Lowly dependent users, on the other hand, may have a low comprehensive preference index but a high angle switching rate, reflecting strong independent composition ability and a varied shooting style. This refined user profile enables the system to provide differentiated functional designs and usage guidance for different types of users.

[0056] Step S106: Design a dynamic ad matching strategy based on the user's composition behavior feature vector. Determine the ad recommendation category by analyzing the matching features of angle diversity index and cue dependence in the user's comprehensive preference index. When the user displays high angle diversity and low cue dependence, recommend professional composition tool ads. When the user displays a single angle but high cue dependence, recommend basic auxiliary function ads. This achieves accurate ad content selection and delivery strategy formulation.

[0057] Based on the user's composition behavior feature vector, the angle switching rate is extracted as an angle diversity indicator, and the comprehensive preference index and dependence level label are extracted as cue dependence indicators. If the angle switching rate is higher than a preset diversity threshold and the dependence level is low dependence, the user is marked as having a professional composition need type. If the angle switching rate is lower than a preset uniformity threshold and the dependence level is high dependence, the user is marked as having a basic auxiliary need type. For each user need type, a mapping relationship is established between advertising content and the need type. The professional composition need type is mapped to a first set of advertisements containing advanced composition theory tutorials, introductions to professional photography equipment, and composition creative tools. The basic auxiliary need type is mapped to a second set of advertisements containing composition beginner guides, automatic composition applications, and simple auxiliary tools. A preset number of candidate advertisements are randomly selected from the corresponding advertisement sets according to the user's need type. The matching degree between each advertisement and the user's characteristics is calculated using these candidate advertisements. The matching degree is equal to the consistency score between the user's need type and the advertisement type multiplied by a preset matching coefficient. The advertisement with the highest matching degree is selected as the recommendation result, achieving precise advertising content selection and delivery strategy formulation.

[0058] Specifically, the user composition behavior feature vector, as a data structure describing user shooting habits, contains quantitative indicators across multiple dimensions.

[0059] In one possible implementation, the angle switching rate reflects how frequently a user changes their perspective during shooting. A photography enthusiast's feature vector shows an angle switching rate of 0.7, indicating that they changed their shooting angle 7 out of 10 times. This high switching rate typically signifies that the user possesses strong compositional awareness and creative ability, and is willing to experiment with different perspectives to achieve the best results.

[0060] Specifically, the dependence level labels are extracted directly from the user classification results obtained in the previous analysis. Low-dependency users are typically characterized by an overall preference index below 0.3, indicating that they rarely use auxiliary tools; high-dependency users have an overall preference index above 0.8, reflecting their strong demand for various auxiliary functions. When the system detects an angle switching rate of 0.7 combined with the low-dependency label, it can be inferred that the user is an experienced photographer with independent composition skills and a diverse creative style.

[0061] It's important to note that the user need type labeling process involves a dual-judgment mechanism. A preset diversity threshold might be set at 0.6; when the angle switching rate exceeds this value, it indicates that the user's shooting style is flexible and varied. A preset uniformity threshold might be 0.2; below this value, it indicates that the user tends to shoot from a fixed angle. This classification method precisely divides users into two main categories: professional composition needs and basic auxiliary needs. The mapping relationship between advertising content and need types reflects the core concept of precision marketing. The first set of advertisements targets professional users, including advanced composition theory tutorials that may cover advanced applications of the golden ratio, dynamic composition techniques, etc.; professional photography equipment introductions recommend high-end cameras, professional lenses, etc.; and composition creative tools may include professional photo editing software, composition analysis applications, etc. The second set of advertisements targets beginners, providing basic knowledge explanations in composition beginner guides, automatic composition applications to help users get started quickly, and simple auxiliary tools to lower the shooting threshold.

[0062] In one embodiment, the random selection of candidate ads is controlled by a preset number to avoid information overload for users. Assuming five candidates are selected from each ad set, the system performs weighted random selection based on factors such as ad update time and historical display frequency to ensure the freshness and diversity of ad content. The core of the matching degree calculation lies in assessing the degree of fit between ad content and user characteristics. The consistency score uses a binary judgment: a score of 1 is given when the user's need type perfectly matches the ad type, and 0 otherwise. The preset matching coefficient may be set differently for different ad categories, with tutorial ads having a higher coefficient and product recommendation ads having a lower coefficient. The matching degree is obtained by multiplying the two scores, and the ad with the highest score is selected for push, realizing a complete closed loop from user behavior analysis to precise ad delivery, improving ad effectiveness and user experience.

[0063] Step S107: Monitor ad click-through rate and conversion behavior, dynamically adjust the weights of the three dimensions of user composition behavior feature vector: shooting angle preference, grid style selection, and tool usage duration. At the same time, optimize ad display strategy by combining the real-time angle detection results of the user's current shooting scene. When it is detected that the user is shooting from above and historical data shows that this angle is a new attempt, increase the display weight of composition tutorial ads and determine the final personalized ad recommendation results.

[0064] A real-time monitoring mechanism records user clicks and subsequent conversions on pushed ads. The click-through rate (CTR) is calculated by comparing the number of clicks to the number of impressions. Purchases, downloads, or registrations after clicking ads are recorded as conversion markers. The CTR is multiplied by the binary result of these conversion markers to obtain an ad performance score. A Bayesian update algorithm processes this ad performance score, using the weights of three dimensions—current shooting angle preference, grid style selection, and duration of tool usage—as prior probabilities. The ad performance score is then used as the input to a likelihood function to calculate updated dimension weights. These updated weights replace the corresponding weights in the original feature vector, forming an adjusted user composition behavior feature vector. Based on this adjusted feature vector, real-time angle data detected by the gyroscope sensor at the time of the user's current shot is obtained. The frequency of each angle in the user's historical shooting data is queried. If the current detected angle is a top-down view and the historical usage frequency of this angle is below a preset threshold, the matching coefficient of composition tutorial ads is increased by a preset increment. The ad matching degree is then recalculated using the adjusted feature vector, and the ad with the highest matching degree is selected as the final personalized ad recommendation result.

[0065] Specifically, the core of the real-time monitoring mechanism lies in accurately capturing user interaction behavior with advertisements.

[0066] In one possible implementation, when the system displays an advertisement for "Professional Composition Theory Tutorial" to a user, it records the timestamp and location of the ad. If the user clicks on the ad within 5 seconds, the system records it as a valid click. The click-through rate (CTR) is calculated cumulatively; for example, if the ad is displayed 100 times and clicked 15 times, the CTR would be 0.15.

[0067] Specifically, the conversion flag determination involves tracking multiple user behaviors. After a user clicks on an ad, the system continues to monitor subsequent behavior: if the user completes a tutorial purchase, the conversion flag is set to 1; if the user only browses and then exits, the conversion flag is set to 0. The ad performance score is obtained by multiplying the click-through rate (CTR) by the conversion flag. In the above example, if there are 3 conversions out of 15 clicks, the average conversion rate is 0.2, then the ad performance score is 0.15 × 0.2 = 0.03. This dual-metric design avoids the problem of simply pursuing CTR while ignoring actual conversion performance. The application of the Bayesian update algorithm in this scenario demonstrates the characteristics of dynamic learning. The user's current feature vector weights are used as prior information, such as a weight of 0.4 for photo angle preference, 0.35 for grid style selection, and 0.25 for tool usage duration. When the ad performance score is 0.03, it indicates that the current recommendation strategy is not ideal, and the algorithm will reduce the weight of the corresponding dimension. The specific calculation process involves using the advertising effectiveness score as observational evidence and updating the probability distribution of the weights of each dimension through the Bayesian formula, resulting in new weight values ​​that may be adjusted to 0.38, 0.33, or 0.29.

[0068] It should be noted that real-time angle detection relies on the device's built-in gyroscope sensor, which can accurately obtain the spatial attitude of the current shooting device. When the sensor detects that the device's tilt angle exceeds 45 degrees and the lens is pointing downwards, it is determined to be a top-down shot. The query of historical usage frequency is achieved by traversing the user's past shooting records. If the user has only used the top-down angle 20 times in the last 1000 shots, the frequency is 0.02, which is lower than the preset threshold of 0.05, and the system determines that this is a new attempt.

[0069] In one embodiment, the weighting mechanism for composition tutorial ads employs a dynamic adjustment strategy. When a new attempt is detected, the system assumes the user is exploring new shooting techniques, and recommending related tutorials at this time has a high acceptance rate. The incremental value of the matching coefficient might be set to 0.2, increasing the original matching coefficient of the composition tutorial ad from 0.5 to 0.7. After recalculating the matching degree of all candidate ads, the composition tutorial ad is likely to become the final recommendation result. This dynamic adjustment mechanism based on real-time feedback forms a closed-loop optimization system. Every user interaction affects subsequent recommendation strategies, and the system continuously learns user preferences to optimize the accuracy of ad recommendations. In particular, the identification and response to new attempts demonstrates the system's keen insight into user learning needs, improving the timeliness and relevance of ad recommendations.

[0070] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for precise ad delivery based on multi-dimensional user profiles, characterized in that, The method includes: Distribution features of top-down, bottom-up, and eye-level angles are extracted from users' historical shooting data. The frequency of use of each angle type over time is calculated to construct a dynamic change vector of user shooting angle preferences. Based on this dynamic change vector, an angle-grid mapping relationship is established: top-down angles are mapped to a rule-of-thirds grid pattern, and eye-level angles are mapped to a golden spiral grid pattern, generating a personalized grid pattern preference rating matrix. Using this personalized grid pattern preference rating matrix, users are categorized into three groups—high-dependency, moderate-dependency, and low-dependency—using a clustering algorithm. The activation frequency and dwell time of the golden section point prompt function for each group are statistically analyzed to form a user composition guide. The system generates a user composition aid usage behavior profile; based on this profile, it determines the user's reliance level on the golden ratio point prompt; it integrates the user's dynamic change vector of shooting angle preferences, the personalized grid style preference rating matrix, and the dwell time in the user composition aid usage behavior profile, and uses a weighted fusion algorithm to calculate a comprehensive preference index, generating a user composition behavior feature vector; it then calculates the angle switching rate in the user composition behavior feature vector as an angle diversity indicator; based on the user composition behavior feature vector, combined with the angle diversity indicator and the golden ratio point prompt reliance level, it determines the ad recommendation category and generates personalized ad recommendation results.

2. The precise advertising method based on multi-dimensional user profiles according to claim 1, characterized in that, The process involves extracting the distribution characteristics of top-down, bottom-up, and eye-level angles from historical user shooting data, calculating the rate of change in the frequency of use of each angle type over time, and constructing a dynamic change vector of user shooting angle preferences, including: The system extracts the shooting timestamp of each photo from the user's historical shooting data, extracts the tilt angle value from the camera gyroscope sensor, and determines the angle as a downward tilt (camera pointing downwards with a tilt angle greater than 45 degrees), an upward tilt (camera pointing upwards with a tilt angle greater than 45 degrees), and a level tilt (tilt angle between ±45 degrees). It then statistically analyzes the shooting frequency of each angle type within a time period to form an angle distribution feature matrix. Based on this matrix, it divides time windows, calculates the rate of change of frequency for each angle type within adjacent time windows, adjusts the weight coefficients for each angle type based on the rate of change, and accumulates the weight coefficient changes to generate a trend change coefficient. Finally, it normalizes the trend change coefficient and the weight coefficients for each angle type to generate a dynamic change vector of user shooting angle preferences, where each element in the vector is the product of the normalized weight coefficient of the angle type and the trend change coefficient.

3. The precise advertising method based on multi-dimensional user profiles according to claim 1, characterized in that, The step of establishing an angle-grid mapping relationship based on the dynamic change vector of the user's shooting angle preference, mapping the top-down angle to a rule-of-thirds grid pattern, and mapping the eye-level angle to a golden spiral grid pattern, generates a personalized grid pattern preference rating matrix, including: Based on the angle type weights and trend coefficients in the dynamic vector of user shooting angle preferences, the mapping strength between the top-down angle and the third-order grid pattern, and the mapping strength between the eye-level angle and the golden spiral grid pattern are calculated. The mapping strength is the product of the angle trend coefficient and the normalized weight coefficient. Based on the mapping strength, the recommendation weight of the third-order grid pattern is adjusted to the mapping strength of the top-down angle, and the matching priority of the golden spiral grid pattern is adjusted. Based on the recommendation weight of the third-order grid pattern and the matching priority of the golden spiral grid pattern, a personalized grid pattern preference scoring matrix is ​​constructed. Each row in the matrix corresponds to a grid pattern, and each column is the preference score for that grid pattern.

4. The precise advertising method based on multi-dimensional user profiles according to claim 1, characterized in that, The personalized grid style preference rating matrix is ​​used to categorize users into three types—high dependency, moderate dependency, and low dependency—using a clustering algorithm. The sum of the three-part grid score and the golden spiral grid score in the personalized grid style preference rating matrix is ​​calculated to generate a total score of the user's dependence on the composition aid tool. The user is labeled as high dependence, moderate dependence, or low dependence based on the total score. Based on the total dependence score, a clustering algorithm is used to further segment the user and generate three user groups: high dependence, moderate dependence, and low dependence.

5. The precise advertising method based on multi-dimensional user profiles according to claim 4, characterized in that, The determination of the user's reliance level on the golden ratio point prompts based on the user's mapping assistance tool usage behavior profile includes: Extract the activation frequency and dwell time from the user's mapping assistance tool usage behavior profile, and label the user's golden section point prompt dependency level as high, medium, or low based on the activation frequency and dwell time; calculate the overall usage rate of the grid assistance tool according to the golden section point prompt dependency level, and determine that high-level users are highly dependent, medium-level users are moderately dependent, and low-level users are low dependent.

6. The precise advertising method based on multi-dimensional user profiles according to claim 1, characterized in that, The comprehensive preference index is calculated using a weighted fusion algorithm by integrating the dynamic change vector of the user's shooting angle preference, the personalized grid style preference rating matrix, and the dwell time in the user's composition assistance tool usage behavior profile. This includes: The angle weight values ​​of the dynamic change vector of the user's photo angle preference, the grid score values ​​of the personalized grid style preference scoring matrix, and the dwell time of the user's composition assistance tool usage profile are normalized to generate standardized feature values. Based on the standardized feature values, a weighted summation method is used to calculate the comprehensive preference index, which is the sum of the products of the standardized feature values ​​of each dimension and the preset weights.

7. The precise advertising method based on multi-dimensional user profiles according to claim 6, characterized in that, The step of determining the ad recommendation category based on the user's mapping behavior feature vector, combined with the angle diversity index and the golden ratio cue dependence level, includes: The angle switching rate and comprehensive preference index are extracted from the user's composition behavior feature vector. Based on the angle switching rate and the golden ratio point prompt dependence level, the user is labeled as either a professional composition need type or a basic auxiliary need type. According to the user's need type, it is mapped to the corresponding ad set, and the ad with the highest matching degree is selected from the ad set as the recommendation result.

8. The precise advertising method based on multi-dimensional user profiles according to claim 7, characterized in that, The generation of personalized advertising recommendation results includes: Record user clicks and conversions on pushed ads, calculate click-through rate and ad performance score; based on the ad performance score, use a Bayesian update algorithm to adjust the dimension weights in the user composition behavior feature vector; combine the user composition behavior feature vector and real-time angle detection data to adjust the matching coefficient of composition tutorial ads, and select the ad with the highest matching degree as the personalized ad recommendation result.

Images