System for dynamically adjusting advertisement delivery weight based on user real-time behavior characteristics

Abstract

The application relates to the technical field of advertisement delivery, and particularly discloses a system for dynamically adjusting advertisement delivery weight based on real-time behavior characteristics of users, which comprises a behavior state space construction module, a user group layering module, a group reference manifold construction module, a behavior deviation vector calculation module and a weight correction vector generation module. The behavior state space construction module is used for constructing a user behavior state space; the user group layering module is used for dividing user layering grid units; the group reference manifold construction module is used for fitting a group reference manifold; the behavior deviation vector calculation module is used for determining a target reference point on the group reference manifold and constructing a behavior deviation vector and extracting a deviation geometric module length; the weight correction vector generation module is used for decomposing the behavior deviation vector into a normal component and a tangent component, taking the direction of the normal component as a weight correction direction, taking the module length of the normal component as a calculation basis of a weight correction amplitude, and generating a weight correction vector; the weight correction execution module is used for correcting an initial weight configuration; and the delivery execution module is used for executing advertisement delivery. The application improves the precision of advertisement delivery weight adjustment.

Description

Technical Field

[0001] This invention relates to the field of advertising delivery technology, and in particular to a system for dynamically adjusting advertising delivery weights based on real-time user behavior characteristics. Background Technology

[0002] In the field of advertising delivery technology, dynamically adjusting ad delivery weights based on user behavior characteristics is a core means of achieving precision marketing. Current technologies typically collect user behavior data such as clicks, browsing, interactions, and conversions, calculate user interest preference weights for various types of ads using statistical analysis methods, and adjust subsequent ad delivery strategies accordingly. With the development of user behavior modeling technology, some existing technologies have begun to incorporate machine learning methods for feature extraction and cluster analysis of user behavior to identify the behavioral patterns of different user groups, thereby achieving differentiated ad weight configurations. Furthermore, the application of edge computing architecture in the advertising delivery field has been publicly reported, used to reduce data processing latency and improve the real-time performance of ad delivery.

[0003] However, existing technologies still suffer from the following fundamental flaws: The adjustment of ad placement weights is essentially limited to independent judgments of the absolute values ​​of individual behaviors, failing to place individual behaviors within a holistic reference frame of the group's behavioral distribution for relative evaluation. This results in an inability to effectively distinguish between "individual behavioral changes within the normal fluctuation range of the group" and "abnormal behaviors that truly deviate from the mainstream behavioral patterns of the group," leading to insufficient precision and frequent erroneous adjustments. Simultaneously, the weight adjustment strategy lacks geometric modeling and quantification methods for the direction of behavioral deviations, failing to map the systematic directional information of user behavior deviations into targeted corrections of ad placement weights. This results in ambiguous directionality of weight corrections, hindering the achievement of differentiated and precise weight allocation between different ad types. The combined effect of these flaws makes it difficult for the accuracy and adaptability of dynamic ad placement weight adjustments to meet actual business needs.

[0004] Therefore, this invention proposes a system for dynamically adjusting the weight of ad delivery based on real-time user behavior characteristics. Summary of the Invention

[0005] This invention provides a system for dynamically adjusting advertising weights based on real-time user behavior characteristics. It constructs a user behavior reference system at the same level through a group baseline manifold, decomposes the individual behavior deviation vector into normal and tangential components, and uses the normal component to accurately drive the retargeting correction of advertising weights. This solves the technical defects of ambiguous weight adjustment direction and frequent misadjustments in the prior art, and realizes accurate and differentiated dynamic adjustment of advertising weights.

[0006] This invention provides a system for dynamically adjusting ad delivery weights based on real-time user behavior characteristics, comprising: The Behavior State Space Construction Module is used to construct a user behavior state space with behavior type as the dimension axis and behavior time coordinates and intensity coordinates as coordinate components. The user behavior state is expressed as the behavior state coordinate points in the user behavior state space. The user group segmentation module is used to divide all users into multiple user segmentation grid units according to preset segmentation attributes and segmentation levels; The group baseline manifold construction module is used to fit and generate a group baseline manifold based on the set of historical behavioral state coordinates of all users within the hierarchical grid cell of the target user. The behavior deviation vector calculation module is used to determine the target reference point on the group's baseline manifold. It constructs the behavior deviation vector with the real-time behavior state coordinates of the target user as the starting point and the target reference point as the ending point. The weight correction vector generation module is used to decompose the behavior deviation vector into the normal component of the normal direction of the group reference manifold and the tangential component of the tangential direction of the group reference manifold. The direction of the normal component is used as the weight correction direction, and the magnitude of the normal component is used as the basis for calculating the weight correction magnitude to generate the weight correction vector. The weight correction execution module is used to correct the initial ad delivery weight configuration based on the weight correction vector and generate the corrected ad delivery weight configuration. The ad delivery execution module is used to execute ad delivery operations based on the revised ad delivery weight configuration.

[0007] Preferably, the behavior state space construction module includes: The behavior type dimension axis definition unit is used to define each behavior type in the preset behavior type set as a dimension axis of the user behavior state space. The preset behavior type set includes click behavior type, browsing duration behavior type, interaction frequency behavior type, and conversion completion behavior type. The time coordinate component mapping unit is used to normalize the occurrence timestamp of the target behavior according to the preset time window length, and use the normalized time value as the first coordinate component of the target behavior on the corresponding behavior type dimension axis. The intensity coordinate component mapping unit is used to convert the original interaction data of the target behavior into the behavior intensity quantization value according to the preset behavior intensity quantization rule, and use the behavior intensity quantization value as the second coordinate component of the target behavior on the corresponding behavior type dimension axis. The coordinate point generation unit is used to combine the first and second coordinate components on the dimension axis of all behavior types of the target user at the target time to generate the behavior state coordinate points of the target user at the target time.

[0008] Preferably, the group baseline manifold construction module includes: The same-layer historical coordinate aggregation unit is used to extract the set of historical behavior state coordinates of all users in the target user's hierarchical grid unit within a preset historical time period from the user database. The manifold structure definition unit is used to define the manifold structure of the group's baseline manifold as a locally linearly embedded manifold, where the embedding dimension of the locally linearly embedded manifold is smaller than the dimension of the user behavior state space. The manifold parameter fitting unit is used to fit the manifold parameters of the local linearly embedded manifold using the set of historical behavior state coordinate points as training data and minimizing the sum of squared projection residuals from each coordinate point in the set of historical behavior state coordinate points to the local linearly embedded manifold as the optimization objective, thereby generating the fitted population baseline manifold.

[0009] Preferably, the behavior deviation vector calculation module includes: The manifold projection calculation unit is used to calculate the projection coordinates of the real-time behavior state coordinates on the group reference manifold based on the local linear reconstruction relationship of the group reference manifold. The projection coordinates are the coordinates on the group reference manifold that are closest to the real-time behavior state coordinates. The historical trajectory positioning unit is used to project each historical behavior state coordinate point onto the group reference manifold surface according to the historical behavior state coordinate point sequence of the target user to obtain a set of historical trajectory projection points on the group reference manifold. The historical projection point in the set of historical trajectory projection points that is closest to the current time is used as the target reference point on the group reference manifold. The deviation vector determination unit is used to construct a behavior deviation vector starting from the target reference point and ending at the real-time behavior state coordinate point.

[0010] Preferably, the weight correction vector generation module includes: The normal direction extraction unit is used to extract the normal direction of the real-time behavior state coordinate point at the projected coordinate point on the group reference manifold, and use the normal direction as the reference direction for weight correction. The normal component decomposition unit is used to project the behavior deviation vector onto the weight correction reference direction to obtain the normal component of the behavior deviation vector in the direction of the group reference manifold normal. The normal magnitude normalization unit is used to compare the magnitude of the normal component with a preset magnitude threshold. When the magnitude of the normal component is less than or equal to the preset magnitude threshold, the weight correction magnitude is set to zero. When the magnitude of the normal component is greater than the preset magnitude threshold, the difference between the magnitude of the normal component and the preset magnitude threshold is used as the weight correction magnitude. The weight correction vector synthesis unit is used to synthesize a weight correction vector by taking the unit direction of the normal component as the weight correction direction component and the weight correction magnitude component as the weight correction magnitude component.

[0011] Preferably, the normal component decomposition unit includes: The tangential component calculation sub-unit is used to subtract the normal component from the behavior deviation vector to obtain the tangential component of the behavior deviation vector in the tangential direction of the population reference manifold; The weight correction direction selection subunit is used to compare the magnitudes of the normal and tangential components. When the magnitude of the normal component is greater than that of the tangential component, the direction of the normal component is output as the weight correction direction. When the magnitude of the normal component is less than or equal to that of the tangential component, it is determined that the real-time behavior state of the target user is within the range of the group reference manifold, and the amplitude is set to zero signal to the normal magnitude normalization unit.

[0012] Preferably, the weight correction execution module includes: The correction direction mapping unit is used to determine the projection component value of the weight correction vector direction on each behavior type dimension axis based on the direction cosine of the weight correction vector in the user behavior state space, and to determine the ad type corresponding to the behavior type dimension axis with the largest projection component value as the ad type to be corrected according to the preset mapping relationship between the behavior type dimension axis and the ad type. The correction calculation unit is used to multiply the weight correction magnitude by a preset magnitude mapping coefficient to generate the correction value for each type of advertisement to be corrected. The preset magnitude mapping coefficient is positively correlated with the placement intensity parameter in the advertiser's placement target parameters. The weight configuration correction unit is used to add the corresponding correction value to the weight components corresponding to each type of advertisement to be corrected in the initial advertisement placement weight configuration, and generate the corrected advertisement placement weight configuration.

[0013] Preferably, it further includes: The actual response feedback collection module is used to collect the actual delivery response data of target users under the corrected ad delivery weight configuration. The actual delivery response data includes the actual click-through rate and the actual conversion rate. The manifold correction module is used to generate a manifold correction signal based on the difference between the actual delivery response data and the predicted manifold behavior response value. The manifold correction signal is then fed back to the population baseline manifold construction module, which uses the manifold correction signal as a constraint to adjust the manifold parameters of the population baseline manifold. The vector correction module generates a vector correction signal based on the difference between the actual delivery response data and the predicted delivery effect response value. The vector correction signal is then fed back to the weight correction vector generation module to update the calculation mapping parameters of the weight correction magnitude. The predicted delivery effect response value is calculated by the preset delivery effect prediction model based on the corrected advertising delivery weight configuration.

[0014] Preferably, the manifold correction module includes: The prediction response calculation unit is used to calculate the predicted response value of the manifold behavior in the reference manifold by taking the projected coordinates of the real-time behavior state coordinates as the reference, offsetting the behavior deviation vector along the projection direction on the manifold tangent plane and the geodesic of the manifold by a preset step size, and then inputting the weight correction vector corresponding to the offset coordinates into the delivery effect prediction model to obtain the predicted response value of the manifold behavior corresponding to the offset coordinates. The correction signal generation unit is used to calculate the difference between the actual delivery response data and the predicted response value of the manifold behavior. When the difference is positive, a first correction signal is generated, which indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is increased. When the difference is negative, a second correction signal is generated, which indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is decreased. The manifold parameter refitting unit is used to perform incremental refitting operations on the manifold parameters of the population reference manifold based on a first correction signal or a second correction signal.

[0015] Preferably, it further includes: An edge computing node cluster consists of multiple edge computing nodes distributed in different geographical locations, with each edge computing node communicating with user terminal devices within its corresponding geographical area. The deviation vector local cache unit is set inside each edge computing node to cache the behavior deviation vector and deviation geometric magnitude of each user in the corresponding geographical area of ​​the edge computing node; The local weight pre-correction unit is set inside each edge computing node. When the communication connection between the edge computing node and the group baseline manifold construction module is interrupted, it performs weight correction vector generation and weight correction execution operations locally on the edge computing node based on the locally cached behavior deviation vector and deviation geometric magnitude to generate a local emergency weight configuration. The local emergency weight configuration is replaced by the latest corrected advertising delivery weight configuration generated by the weight correction execution module after the communication connection is restored.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: Existing technologies rely on independent judgments of the absolute values ​​of individual behaviors when adjusting advertising weights. The direction of weight correction is ambiguous and cannot distinguish between normal behavioral fluctuations and truly abnormal behaviors that deviate from the mainstream of the group, resulting in serious deficiencies in the accuracy and directionality of weight adjustments. This invention constructs a user behavior state space, expressing individual behaviors as high-dimensional geometric coordinate points. It then uses machine learning methods to introduce a group baseline manifold as a low-dimensional embedding reference system for the distribution of user behaviors at the same level. Using the most recent projection point of the target user's historical behavior on the manifold as the target reference point, a behavior deviation vector from the target reference point to the real-time behavior state coordinate point is constructed. This vector is then decomposed along the manifold normal direction, transforming the behavior deviation into a geometric vector with clear directional and magnitude meanings. The direction of the normal component precisely indicates the direction of the advertising type that needs correction, and the magnitude of the normal component precisely quantifies the weight correction magnitude. This solves the technical problems of ambiguous attribution of behavioral changes and lack of directionality in weight correction, enabling the adjustment of advertising weights to have a precise, geometric, and directional driving mechanism, thus improving the accuracy and business adaptability of dynamic adjustments to advertising weights. This invention achieves refined dynamic modeling of user group behavior distribution through a group baseline manifold, constructing a group behavior reference benchmark with clear geometric meaning for weight adjustment; the directional correction mechanism based on normal components can directly transform behavioral deviation information into precise weight adjustment actions, effectively supporting the refined strategy optimization of advertising.

[0017] This invention achieves accurate determination of the validity of behavioral deviations by comparing the normal and tangential component magnitudes after decomposition of the behavioral deviation vector: when a user's behavioral state migrates along the surface of the group's baseline manifold and the tangential component magnitude dominates, it is determined to be a normal evolution of the user within the group's behavioral pattern, and no weight correction is triggered, effectively avoiding frequent erroneous adjustments; when a user's behavioral state deviates from the surface of the group's baseline manifold and the normal component magnitude is significantly higher than the tangential component magnitude, it is determined to be a genuine abnormal deviation, and weight correction is initiated. This mechanism solves the technical problem of existing technologies being unable to distinguish between "individual behavioral changes within the normal fluctuation range of the group" and "individual behavioral abnormalities that truly deviate from the mainstream behavioral pattern of the group."

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall module connection of the system for dynamically adjusting the advertising delivery weight based on real-time user behavior characteristics in an embodiment of the present invention. Figure 2 This is a schematic diagram of the internal unit connections of the behavior state space construction module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the internal unit connections of the group reference manifold construction module in an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] like Figure 1 As shown, this invention provides an embodiment of a system for dynamically adjusting ad delivery weights based on real-time user behavior characteristics, including: The Behavior State Space Construction Module is used to construct a user behavior state space with behavior type as the dimension axis and behavior time coordinates and intensity coordinates as coordinate components. The user behavior state is expressed as the behavior state coordinate points in the user behavior state space. The user group segmentation module is used to divide all users into multiple user segmentation grid units according to preset segmentation attributes and segmentation levels; The group baseline manifold construction module is used to fit and generate a group baseline manifold based on the set of historical behavioral state coordinates of all users within the hierarchical grid cell of the target user. The behavior deviation vector calculation module is used to determine the target reference point on the group's baseline manifold, and construct the behavior deviation vector with the target reference point as the starting point and the real-time behavior state coordinate point of the target user as the ending point. The weight correction vector generation module is used to decompose the behavior deviation vector into the normal component of the normal direction of the group reference manifold and the tangential component of the tangential direction of the group reference manifold. The direction of the normal component is used as the weight correction direction, and the magnitude of the normal component is used as the basis for calculating the weight correction magnitude to generate the weight correction vector. The weight correction execution module is used to correct the initial ad delivery weight configuration based on the weight correction vector and generate the corrected ad delivery weight configuration. The ad delivery execution module is used to execute ad delivery operations based on the revised ad delivery weight configuration.

[0023] In this embodiment, the behavior types include click behavior type, browsing duration behavior type, interaction frequency behavior type, and conversion completion behavior type.

[0024] In this embodiment, the behavior time coordinate refers to the value obtained by normalizing the time of occurrence of the user behavior according to the preset time window length. The value range of the behavior time coordinate is between 0 and 1, representing the relative occurrence time of the behavior within the preset time window.

[0025] In this embodiment, the intensity coordinate refers to the value obtained after converting the original interaction data of user behavior according to the preset behavior intensity quantification rules. The value of the intensity coordinate reflects the degree of influence of the behavior on the adjustment of the advertising placement weight. Different behavior types correspond to different intensity coordinate calculation methods.

[0026] In this embodiment, the intensity coordinate of the browsing duration behavior type is equal to the ratio of the browsing duration of a single advertisement by the target user within a preset time window to the length of the preset time window.

[0027] In this embodiment, the user behavior state space is a high-dimensional space with behavior type as the dimension axis and behavior time coordinate and intensity coordinate as coordinate components. The number of dimensions of this space is equal to the number of behavior types multiplied by 2. The coordinate value on each dimension axis is a two-dimensional vector. The first component of the two-dimensional vector is the behavior time coordinate, and the second component of the two-dimensional vector is the intensity coordinate.

[0028] In this embodiment, the user behavior state refers to the coordinate point in the user behavior state space used to characterize the user's behavior characteristics at a specific moment. The user behavior state includes the behavior time coordinate and intensity coordinate on the dimension axis of all behavior types at that moment.

[0029] In this embodiment, the user behavior state coordinate point is a numerical representation of the user behavior state in the user behavior state space, and each user behavior state coordinate point corresponds to the user's behavior state at a certain moment.

[0030] In this embodiment, the preset hierarchical attributes include geographic region attributes, device type attributes, and active time period attributes. The hierarchical levels of geographic region attributes include national level, provincial level, and city level. The hierarchical levels of device type attributes include mobile terminal device level and fixed terminal device level. The hierarchical level of active time period attributes is determined based on the time period distribution density of user historical behavior data.

[0031] In this embodiment, the user hierarchical grid unit is a user group group obtained by combining all users according to preset hierarchical attributes and hierarchical levels. A user hierarchical grid unit is jointly defined by the geographic region hierarchical node, device type hierarchical node and active time period hierarchical node to which the user belongs. Users in the same user hierarchical grid unit have the same geographic region hierarchical level, device type hierarchical level and active time period hierarchical level.

[0032] In this embodiment, the target user hierarchical grid cell refers to the user hierarchical grid cell to which the target user belongs, and is used to determine the same-layer user group that has the same hierarchical attributes as the target user.

[0033] In this embodiment, the set of historical behavior state coordinates of a user refers to the summary set of the behavior state coordinates of all users in the same user hierarchical grid cell at various times within a preset historical period. This set includes the behavior time coordinates and intensity coordinates of each user in the user hierarchical grid cell at each preset historical time.

[0034] In this embodiment, the group baseline manifold is a low-dimensional embedding surface generated by processing the set of historical behavior state coordinate points of all users within the hierarchical grid cell of the target user through a manifold fitting operation. The dimension of this low-dimensional embedding surface is smaller than the dimension of the user behavior state space, and it represents the overall distribution trend of user behavior states within the same hierarchical grid cell in the user behavior state space. The manifold fitting operation uses a locally linear embedding method, which is constructed as follows: each coordinate point in the set of historical behavior state coordinate points is represented as a linear weighted combination of its nearest neighbor coordinate points. The linear weights are solved by minimizing the reconstruction error, and then the coordinate points are reconstructed in the low-dimensional space based on the linear weights, so that the reconstruction relationship between coordinate points in the low-dimensional space is consistent with the reconstruction relationship in the high-dimensional space. The resulting low-dimensional embedding surface is the group baseline manifold. The input of the locally linear embedding method is the set of historical behavior state coordinate points, and the output is the group baseline manifold.

[0035] In this embodiment, the deviation geometric modulus refers to the length of the behavior deviation vector, which is obtained by calculating the Euclidean norm of the behavior deviation vector. The deviation geometric modulus characterizes the degree to which the real-time behavior state of the target user deviates from the group's baseline manifold.

[0036] In this embodiment, the weight correction direction is the direction of the normal component of the behavior deviation vector in the direction of the normal of the group reference manifold, the weight correction magnitude is calculated based on the magnitude of the normal component, and the weight correction vector is a vector with the weight correction direction as its direction and the weight correction magnitude as its magnitude. The weight correction vector contains both direction information and magnitude information.

[0037] In this embodiment, the initial ad placement weight configuration refers to the set of weight values ​​that the system pre-sets for each ad type according to the advertiser's placement target parameters before analyzing the real-time behavioral characteristics of the target users. The initial ad placement weight configuration includes the initial weight values ​​for each ad type.

[0038] In this embodiment, the weight correction execution module determines the ad placement weight component to be corrected according to the direction of the weight correction vector, determines the correction value of each ad placement weight component to be corrected according to the magnitude of the weight correction vector, adds the corresponding correction value to the weight component of each ad type to be corrected in the initial ad placement weight configuration, and generates the corrected ad placement weight configuration.

[0039] In this embodiment, the delivery execution module obtains the corrected ad delivery weight configuration, determines the delivery priority and delivery ratio of each ad type according to the weight value of each ad type in the corrected ad delivery weight configuration, and pushes the ad content to the target user's terminal device.

[0040] like Figure 2 As shown, in another embodiment of the present invention, the behavior state space construction module includes: The behavior type dimension axis definition unit is used to define each behavior type in the preset behavior type set as a dimension axis of the user behavior state space. The preset behavior type set includes click behavior type, browsing duration behavior type, interaction frequency behavior type, and conversion completion behavior type. The time coordinate component mapping unit is used to normalize the occurrence timestamp of the target behavior according to the preset time window length, and use the normalized time value as the first coordinate component of the target behavior on the corresponding behavior type dimension axis. The intensity coordinate component mapping unit is used to convert the original interaction data of the target behavior into the behavior intensity quantization value according to the preset behavior intensity quantization rule, and use the behavior intensity quantization value as the second coordinate component of the target behavior on the corresponding behavior type dimension axis. The coordinate point generation unit is used to combine the first and second coordinate components on the dimension axis of all behavior types of the target user at the target time to generate the behavior state coordinate points of the target user at the target time.

[0041] In this embodiment, each dimension axis of the user behavior state space corresponds to a behavior type, and the number of dimension axes is equal to the number of behavior types.

[0042] In this embodiment, click behavior type refers to the category of user behavior that involves clicking on advertising content; browsing duration behavior type refers to the category of user behavior that involves the length of time spent watching advertising content; interaction frequency behavior type refers to the category of user behavior that involves liking, commenting, collecting, sharing, or performing other interactive operations on advertising content; and conversion completion behavior type refers to the category of user behavior that involves completing conversion goals such as purchasing, registering, downloading, or filling out forms.

[0043] In this embodiment, the preset time window length refers to the time interval length that the system pre-sets for normalizing timestamps. The preset time window length is related to the timeliness requirements of the advertising placement scenario.

[0044] In this embodiment, the first coordinate component is a value obtained by normalizing the timestamp of the target behavior within a preset time window length. The normalization process is as follows: the difference between the timestamp of the target behavior and the start time of the preset time window is used as the numerator, and the preset time window length is used as the denominator to calculate the ratio between the two. This ratio is the normalized time value, and the normalized time value is used as the first coordinate component of the target behavior on the corresponding behavior type dimension axis.

[0045] In this embodiment, the preset behavior intensity quantification rule refers to the conversion rule set in advance by the system to convert the original interaction data of the target behavior into behavior intensity quantification value. Different behavior types correspond to different original interaction data formats and different conversion methods.

[0046] In this embodiment, the second coordinate component is the behavior intensity quantization value obtained by transforming the original interaction data of the target behavior according to the preset behavior intensity quantization rule. The behavior intensity quantization value is used as the second coordinate component of the target behavior on the corresponding behavior type dimension axis.

[0047] In this embodiment, the target user's behavior state coordinate point at the target time is a coordinate point obtained by combining the first and second coordinate components on the dimension axis of all behavior types of the target user at the target time. The number of dimensions of this behavior state coordinate point is equal to the number of behavior types multiplied by 2.

[0048] like Figure 3 As shown, in another embodiment of the invention, the population baseline manifold construction module includes: The same-layer historical coordinate aggregation unit is used to extract the set of historical behavior state coordinates of all users in the target user's hierarchical grid unit within a preset historical time period from the user database. The manifold structure definition unit is used to define the manifold structure of the group's baseline manifold as a locally linearly embedded manifold, where the embedding dimension of the locally linearly embedded manifold is smaller than the dimension of the user behavior state space. The manifold parameter fitting unit is used to fit the manifold parameters of the local linearly embedded manifold using the set of historical behavior state coordinate points as training data and minimizing the sum of squared projection residuals from each coordinate point in the set of historical behavior state coordinate points to the local linearly embedded manifold as the optimization objective, thereby generating the fitted population baseline manifold.

[0049] In this embodiment, the user database refers to the data storage unit used by the system to store all user data. The user database contains each user's user identifier, hierarchical attribute information, historical behavior data, and historical behavior status coordinates.

[0050] In this embodiment, the preset historical time period refers to the time interval pre-set by the system for extracting the set of historical behavior state coordinate points. The length of the preset historical time period is determined according to the historical data accumulation cycle of the advertising scenario.

[0051] In this embodiment, the same-layer historical coordinate aggregation unit uses the target user hierarchical grid unit identifier as the retrieval condition to read all historical behavior data of all users belonging to the target user hierarchical grid unit within a preset historical time period from the user database, and converts all the read historical behavior data into a set of historical behavior state coordinate points. The first coordinate component and the second coordinate component on the dimension axis of all behavior types of the same user at the same time constitute a historical behavior state coordinate point of the user at that time.

[0052] In this embodiment, the locally linearly embedded manifold refers to a low-dimensional embedded surface constructed in the user behavior state space using the locally linearly embedded method. The embedding dimension of the locally linearly embedded manifold is smaller than the dimension of the user behavior state space. The value of the embedding dimension is determined based on the number of dimensions of the user behavior state space and the number of users within the target user hierarchical grid cell.

[0053] In this embodiment, the embedding dimension refers to the dimension value of the local linear embedding manifold in the user behavior state space. The embedding dimension is smaller than the dimension of the user behavior state space. The value of the embedding dimension minimizes the sum of squared projection residuals of the coordinate points in the set of historical behavior state coordinate points onto the local linear embedding manifold.

[0054] In this embodiment, the projection residual of each coordinate point in the set of historical behavior state coordinate points to the local linear embedded manifold refers to the shortest distance from each historical behavior state coordinate point to the local linear embedded manifold. This shortest distance reflects the degree of deviation of the historical behavior state coordinate point from the local linear embedded manifold.

[0055] In this embodiment, the manifold parameter fitting unit represents each coordinate point in the historical behavior state coordinate point set as a linear weighted combination of its nearest neighbor coordinate points. It solves for the linear weights by minimizing the reconstruction error, and then reconstructs the coordinate points in a low-dimensional space based on these linear weights, ensuring that the reconstruction relationship between coordinate points in the low-dimensional space is consistent with the reconstruction relationship in the high-dimensional space. The manifold parameter fitting unit optimizes by minimizing the sum of squared projection residuals from each coordinate point in the historical behavior state coordinate point set to the locally linearly embedded manifold, fitting the manifold parameters of the locally linearly embedded manifold to generate the fitted population baseline manifold. The input to the manifold parameter fitting unit is the historical behavior state coordinate point set, and the output is the population baseline manifold.

[0056] In this embodiment, the number of nearest neighbors K in the local linear embedding algorithm ranges from 5 to 15. The embedding dimension is determined using the residual variance inflection point method, and the value of the embedding dimension is set to one-third to one-half of the original behavioral state space dimension.

[0057] In another embodiment of the present invention, the behavior deviation vector calculation module includes: The manifold projection calculation unit is used to calculate the projection coordinates of the real-time behavior state coordinates on the group reference manifold based on the local linear reconstruction relationship of the group reference manifold. The projection coordinates are the coordinates on the group reference manifold that are closest to the real-time behavior state coordinates. The historical trajectory positioning unit is used to project each historical behavior state coordinate point onto the group reference manifold surface according to the historical behavior state coordinate point sequence of the target user to obtain a set of historical trajectory projection points on the group reference manifold. The historical projection point in the set of historical trajectory projection points that is closest to the current time is used as the target reference point on the group reference manifold. The deviation vector determination unit is used to construct a behavior deviation vector starting from the target reference point and ending at the real-time behavior state coordinate point.

[0058] In this embodiment, the real-time behavior state coordinate point refers to the behavior state coordinate point generated by combining the first coordinate component and the second coordinate component of the target user on each behavior type dimension axis at the current moment. The real-time behavior state coordinate point represents the behavior state of the target user at the current moment.

[0059] In this embodiment, the projection coordinate point refers to the coordinate point obtained by vertically projecting the real-time behavior state coordinate point onto the surface of the group reference manifold. The projection coordinate point is the coordinate point on the group reference manifold that is closest to the real-time behavior state coordinate point. The projection coordinate point represents the mapping position of the target user's real-time behavior state on the group reference manifold.

[0060] In this embodiment, the method for calculating the projected coordinate points is as follows: For real-time behavior state coordinate points, the K nearest neighbor coordinate points of the real-time behavior state coordinate points in the original high-dimensional space are matched in the locally linearly embedded manifold of the group reference manifold. The corresponding coordinates of the real-time behavior state coordinate points in the low-dimensional embedding space are calculated using the locally reconstructed weight matrix stored in the locally linearly embedded method. Then, the low-dimensional embedded coordinates are reconstructed back to the original high-dimensional space through locally linear inverse mapping to obtain the projected coordinate points.

[0061] In this embodiment, the behavior deviation vector is a vector constructed with the target reference point as the starting point and the real-time behavior state coordinate point as the ending point. The behavior deviation vector includes a normal component and a tangential component. The normal component is along the normal direction of the group reference manifold at the projected coordinate point, representing the degree to which the real-time behavior state deviates from the surface of the group reference manifold. The tangential component is along the tangent of the group reference manifold at the projected coordinate point, representing the degree to which the real-time behavior state slips along the surface of the group reference manifold.

[0062] In this embodiment, the deviation geometric modulus is equal to the length of the behavior deviation vector. The deviation geometric modulus is a non-negative value. The larger the value of the deviation geometric modulus, the greater the degree to which the real-time behavior state of the target user deviates from the group's baseline manifold.

[0063] In this embodiment, the target reference point refers to the most recent projection point of the target user's historical behavior trajectory onto the group's baseline manifold. Specifically, each coordinate point in the target user's historical behavior state coordinate point sequence within a preset historical time period is projected onto the surface of the group's baseline manifold to obtain a sequence of historical trajectory projection points. The projection point in this sequence that is temporally closest to the current moment is taken as the target reference point. The target reference point characterizes the mapping position of the target user's most recent historical behavior state on the group's baseline manifold and serves as a reference point for measuring the degree of deviation from the current behavior state.

[0064] The target reference point represents the normal baseline position of the target user's recent behavior on the group baseline manifold. The behavior deviation vector takes this group baseline position as a reference and can simultaneously reflect the degree of evolution of the target user's behavior along the group pattern and the degree of deviation from the group pattern.

[0065] In another embodiment of the present invention, the weight correction vector generation module includes: The normal direction extraction unit is used to extract the normal direction of the real-time behavior state coordinate point at the projected coordinate point on the group reference manifold, and use the normal direction as the reference direction for weight correction. The normal component decomposition unit is used to project the behavior deviation vector onto the weight correction reference direction to obtain the normal component of the behavior deviation vector in the direction of the group reference manifold normal. The normal magnitude normalization unit is used to compare the magnitude of the normal component with a preset magnitude threshold. When the magnitude of the normal component is less than or equal to the preset magnitude threshold, the weight correction magnitude is set to zero. When the magnitude of the normal component is greater than the preset magnitude threshold, the difference between the magnitude of the normal component and the preset magnitude threshold is used as the weight correction magnitude. The weight correction vector synthesis unit is used to synthesize a weight correction vector by taking the unit direction of the normal component as the weight correction direction component and the weight correction magnitude component as the weight correction magnitude component.

[0066] In this embodiment, the normal direction refers to the direction perpendicular to the surface of the group reference manifold at the projected coordinate point on the group reference manifold. There are two normal directions, which are opposite to each other. The normal direction represents the local geometric orientation of the group reference manifold at the projected coordinate point. The normal direction extraction unit extracts the normal direction and uses the normal direction as the weight correction reference direction.

[0067] In this embodiment, the normal component refers to the component vector obtained by projecting the behavior deviation vector onto the weight correction reference direction. The magnitude of the normal component is equal to the projection length of the behavior deviation vector onto the weight correction reference direction. The direction of the normal component is the same as or opposite to the weight correction reference direction. The normal component represents the component of the behavior deviation vector that is perpendicular to the group reference manifold surface.

[0068] In this embodiment, the preset modulus threshold is a critical value set in advance by the system to determine whether the modulus of the normal component triggers weight correction. The value of the preset modulus threshold is determined according to the requirements of the advertising scenario on the sensitivity to behavioral deviation.

[0069] In this embodiment, when the magnitude of the normal component is less than or equal to a preset magnitude threshold, it is determined that the degree to which the real-time behavior state of the target user deviates from the group's baseline manifold has not reached the level that requires weight correction, and the weight correction magnitude is set to zero. At this time, the weight correction operation is not triggered.

[0070] In this embodiment, when the magnitude of the normal component is greater than the preset magnitude threshold, it is determined that the real-time behavior state of the target user deviates from the group's baseline manifold to the point where the weight needs to be corrected. The difference between the magnitude of the normal component and the preset magnitude threshold is used as the weight correction magnitude. The larger the value of the weight correction magnitude, the greater the weight correction intensity that needs to be performed.

[0071] In this embodiment, the direction of the weight correction direction component is the unit direction of the normal component, and the value of the weight correction magnitude component is equal to the difference obtained by subtracting the preset magnitude threshold from the magnitude of the normal component. The weight correction vector synthesis unit multiplies the weight correction direction component and the weight correction magnitude component to generate a weight correction vector, which contains weight correction direction information and weight correction magnitude information.

[0072] In another embodiment of the present invention, the normal component decomposition unit includes: The tangential component calculation sub-unit is used to subtract the normal component from the behavior deviation vector to obtain the tangential component of the behavior deviation vector in the tangential direction of the population reference manifold; The weight correction direction selection subunit is used to compare the magnitudes of the normal and tangential components. When the magnitude of the normal component is greater than that of the tangential component, the direction of the normal component is output as the weight correction direction. When the magnitude of the normal component is less than or equal to that of the tangential component, it is determined that the real-time behavior state of the target user is within the range of the group reference manifold, and the amplitude is set to zero signal to the normal magnitude normalization unit.

[0073] In this embodiment, the tangential component calculation subunit subtracts the normal component from the behavior deviation vector to obtain the tangential component of the behavior deviation vector in the tangential direction of the group reference manifold. The direction of the tangential component is parallel to the tangential direction of the group reference manifold at the projected coordinate point. The tangential component represents the component of the behavior deviation vector along the surface of the group reference manifold.

[0074] In this embodiment, the weight correction direction selection subunit extracts the magnitude of the normal component and the magnitude of the tangential component, compares the magnitude of the normal component with the magnitude of the tangential component, and when the magnitude of the normal component is greater than the magnitude of the tangential component, it is determined that the component perpendicular to the group reference manifold in the behavior deviation vector is dominant, and the real-time behavior state of the target user has truly deviated from the group reference. The direction of the normal component is then used as the weight correction direction.

[0075] In this embodiment, when the magnitude of the normal component is less than or equal to the magnitude of the tangential component, the weight correction direction selection subunit determines that the component along the surface of the group reference manifold dominates in the behavior deviation vector. The real-time behavior state of the target user only moves tangentially within the scope of the group reference manifold surface and does not actually deviate from the group reference. Therefore, the weight correction magnitude is set to zero and the weight correction operation is not triggered.

[0076] In another embodiment of the present invention, the weight correction execution module includes: The correction direction mapping unit is used to determine the projection component value of the weight correction vector direction on each behavior type dimension axis based on the direction cosine of the weight correction vector in the user behavior state space, and to determine the ad type corresponding to the behavior type dimension axis with the largest projection component value as the ad type to be corrected according to the preset mapping relationship between the behavior type dimension axis and the ad type. The correction calculation unit is used to multiply the weight correction magnitude by a preset magnitude mapping coefficient to generate the correction value for each type of advertisement to be corrected. The preset magnitude mapping coefficient is positively correlated with the placement intensity parameter in the advertiser's placement target parameters. The weight configuration correction unit is used to add the corresponding correction value to the weight components corresponding to each type of advertisement to be corrected in the initial advertisement placement weight configuration, and generate the corrected advertisement placement weight configuration.

[0077] In this embodiment, the direction cosine refers to the cosine value of the angle between the weight correction direction component and each dimension axis of the user behavior state space. The value of the direction cosine is between -1 and +1, and the value of the direction cosine reflects the projection ratio of the weight correction direction component on each behavior type dimension axis.

[0078] In this embodiment, the comprehensive projection component value is determined as follows: For each behavior type, the sum of the squares of the cosine components of the time coordinate direction and the cosine components of the intensity coordinate direction corresponding to that behavior type is taken as the comprehensive projection component value of that behavior type. The comprehensive projection component values ​​of all behavior types are compared, and the advertisement type corresponding to the behavior type with the largest comprehensive projection component value is determined as the advertisement type to be corrected.

[0079] In this embodiment, the projection component value is the value obtained by multiplying the magnitude of the weight correction direction component by the direction cosine of the weight correction direction component on each behavior type dimension axis. The projection component value characterizes the correction intensity distribution of the weight correction direction component on each behavior type dimension axis. The larger the projection component value, the higher the correction ratio that the behavior type dimension axis needs to bear.

[0080] In this embodiment, the correction direction mapping unit compares the projection component values ​​on each behavior type dimension axis, and determines the ad type corresponding to the behavior type dimension axis with the largest projection component value as the ad type to be corrected. The ad type to be corrected is the target ad type that needs to be weighted.

[0081] In this embodiment, the weight correction magnitude component is equal to the difference between the magnitude of the normal component and the preset magnitude threshold. The weight correction magnitude component is a non-negative value and represents the total intensity of the weight correction to be performed.

[0082] In this embodiment, the preset amplitude mapping coefficient refers to the conversion coefficient that the system pre-sets to convert the weight correction amplitude component into the correction value of each type of advertisement to be corrected. The value of the preset amplitude mapping coefficient is determined according to the overall control requirements of the correction intensity in the advertising scenario.

[0083] In this embodiment, the correction value is the value obtained by multiplying the weight correction amplitude component by a preset amplitude mapping coefficient. The correction value represents the amount of weight adjustment that needs to be performed for each type of advertisement to be corrected.

[0084] In this embodiment, the advertiser's target parameters refer to the set of configuration parameters set by the advertiser to control the advertising strategy, and the advertiser's target parameters include the targeting intensity parameter.

[0085] In this embodiment, the delivery intensity parameter refers to the configuration parameter set by the advertiser to control the overall intensity of the ad delivery. The larger the value of the delivery intensity parameter, the more weight the advertiser wants the delivery system to make in response to changes in behavior.

[0086] In this embodiment, the preset amplitude mapping coefficient is positively correlated with the delivery intensity parameter. When the delivery intensity parameter increases, the preset amplitude mapping coefficient increases accordingly, and when the delivery intensity parameter decreases, the preset amplitude mapping coefficient decreases accordingly.

[0087] In this embodiment, the weight component corresponding to each type of advertisement to be corrected refers to the weight value corresponding to the type of advertisement to be corrected in the initial advertisement placement weight configuration. The initial advertisement placement weight configuration is that each type of advertisement is configured with a weight component.

[0088] In this embodiment, the weight configuration correction unit adds a corresponding correction value to the weight components corresponding to each ad type to be corrected in the initial ad placement weight configuration to generate the corrected ad placement weight configuration. When the correction value is positive, the corresponding weight component increases; when the correction value is negative, the corresponding weight component decreases.

[0089] In another embodiment of the invention, it further includes: The actual response feedback collection module is used to collect the actual delivery response data of target users under the corrected ad delivery weight configuration. The actual delivery response data includes the actual click-through rate and the actual conversion rate. The manifold correction module is used to generate a manifold correction signal based on the difference between the actual delivery response data and the predicted manifold behavior response value. The manifold correction signal is then fed back to the population baseline manifold construction module, which uses the manifold correction signal as a constraint to adjust the manifold parameters of the population baseline manifold. The vector correction module generates a vector correction signal based on the difference between the actual delivery response data and the predicted delivery effect response value. The vector correction signal is then fed back to the weight correction vector generation module to update the calculation mapping parameters of the weight correction magnitude. The predicted delivery effect response value is calculated by the preset delivery effect prediction model based on the corrected advertising delivery weight configuration.

[0090] In this embodiment, after the ad delivery execution module executes the ad delivery operation according to the modified ad delivery weight configuration, the actual response feedback collection module collects the actual delivery response data generated by the target user under the delivery configuration. The actual delivery response data reflects the actual behavioral feedback of the target user to the modified ad delivery weight configuration.

[0091] In this embodiment, the actual click-through rate is equal to the total number of times the target user clicks on the advertised ad under the modified ad placement weight configuration, divided by the total number of times the advertised ad is displayed. The actual conversion rate is equal to the total number of times the target user completes the conversion target under the modified ad placement weight configuration, divided by the total number of times the advertised ad is displayed.

[0092] In this embodiment, the predicted response value of the advertising effect refers to the expected advertising response data of the target user pre-calculated by the weight correction vector generation module when generating the weight correction vector, based on the corrected advertising weight configuration. The predicted response value of the advertising effect includes the predicted click-through rate and the predicted conversion rate.

[0093] In this embodiment, the predicted response value of the campaign performance is calculated by the campaign performance prediction model. The campaign performance prediction model employs a deep neural network structure, comprising an input layer, multiple hidden layers, and an output layer. The input layer receives the weight correction vector corresponding to the behavioral state coordinate points, and the output layer outputs the predicted click-through rate (CTR) and predicted conversion rate. The training data for the campaign performance prediction model consists of historical behavioral state coordinate points and their corresponding historical actual CTR and historical actual conversion rates. During training, the mean squared error between the predicted CTR and the actual CTR, plus the mean squared error between the predicted conversion rate and the actual conversion rate, is used as the total loss function, and the network parameters are adjusted through a backpropagation algorithm.

[0094] In this embodiment, the difference is equal to the actual click-through rate in the actual delivery response data minus the predicted click-through rate in the delivery effect prediction response value, and the actual conversion rate in the actual delivery response data minus the predicted conversion rate in the delivery effect prediction response value. The difference reflects the degree of deviation between the actual delivery effect and the expected delivery effect after the correction of the ad delivery weight configuration.

[0095] In this embodiment, the manifold correction module generates a manifold correction signal based on the difference. The manifold correction signal contains information about the sign of the difference and the magnitude of the difference. The sign of the manifold correction signal indicates the adjustment direction of the group reference manifold, and the magnitude of the manifold correction signal indicates the adjustment range of the group reference manifold.

[0096] In this embodiment, the manifold correction module feeds back the manifold correction signal to the population baseline manifold construction module. The population baseline manifold construction module uses the manifold correction signal as a constraint to perform an incremental refit operation on the manifold parameters of the population baseline manifold to update them. The optimization objective of the incremental refit operation is to enable the population baseline manifold to reduce the difference between the actual delivery response data and the predicted delivery effect response value.

[0097] In this embodiment, the vector correction module generates a vector correction signal based on the difference. The vector correction signal contains information about the sign of the difference and the magnitude of the difference. The sign of the vector correction signal indicates the adjustment direction of the weight correction magnitude calculation mapping parameter, and the magnitude of the vector correction signal indicates the adjustment magnitude of the weight correction magnitude calculation mapping parameter.

[0098] In this embodiment, the vector correction module feeds back the vector correction signal to the weight correction vector generation module. The weight correction vector generation module updates the calculation mapping parameters of the weight correction magnitude according to the vector correction signal. The updated calculation mapping parameters are used for subsequent weight correction magnitude calculation for the target user.

[0099] In another embodiment of the present invention, the manifold correction module includes: The prediction response calculation unit is used to calculate the predicted response value of the manifold behavior in the reference manifold by taking the projected coordinates of the real-time behavior state coordinates as the reference, offsetting the behavior deviation vector along the projection direction on the manifold tangent plane and the geodesic of the manifold by a preset step size, and then inputting the weight correction vector corresponding to the offset coordinates into the delivery effect prediction model to obtain the predicted response value of the manifold behavior corresponding to the offset coordinates. The correction signal generation unit is used to calculate the difference between the actual delivery response data and the predicted response value of the manifold behavior. When the difference is positive, a first correction signal is generated, which indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is increased. When the difference is negative, a second correction signal is generated, which indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is decreased. The manifold parameter refitting unit is used to perform incremental refitting operations on the manifold parameters of the population reference manifold based on a first correction signal or a second correction signal.

[0100] In this embodiment, the preset step size refers to the distance value that the system pre-sets for offsetting on the surface of the group reference manifold. The value of the preset step size is determined according to the scale range of the group reference manifold and the accuracy requirements of the predicted response calculation.

[0101] In this embodiment, the prediction response calculation unit starts from the projected coordinates of the real-time behavior state coordinates in the group baseline manifold, moves along the projection direction of the behavior deviation vector on the tangent plane of the manifold and offsets along the geodesic line of the manifold by a preset step size to reach a new coordinate point. This new coordinate point is used as the offset coordinate point. The weight correction vector corresponding to the offset coordinate point is input into the campaign effect prediction model. The campaign effect prediction model outputs the manifold behavior prediction response value corresponding to the offset coordinate point. The manifold behavior prediction response value includes the predicted click-through rate and the predicted conversion rate.

[0102] In this embodiment, the correction signal generation unit calculates the difference between the actual delivery response data and the manifold behavior prediction response value. The actual delivery response data includes the actual click-through rate and the actual conversion rate, and the manifold behavior prediction response value includes the predicted click-through rate and the predicted conversion rate. The difference is the actual click-through rate minus the predicted click-through rate and the actual conversion rate minus the predicted conversion rate.

[0103] In this embodiment, when the difference is positive, it indicates that the actual delivery response data is higher than the manifold behavior prediction response value. That is, the target user has a higher response performance at the position after the group reference manifold is offset by a preset step size along the tangential projection direction of the behavior deviation vector. The correction signal generation unit generates a first correction signal. The first correction signal indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is increased, so that the group reference manifold bends along the tangential projection direction of the behavior deviation vector, and the manifold surface in this direction moves closer to the position with a higher actual response.

[0104] In this embodiment, when the difference is negative, it indicates that the actual delivery response data is lower than the manifold behavior prediction response value. That is, the target user has a lower response performance at the position after the group reference manifold is offset by a preset step size along the tangential projection direction of the behavior deviation vector. The correction signal generation unit generates a second correction signal. The second correction signal indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector should be reduced, so that the group reference manifold tends to be flat along the tangential projection direction of the behavior deviation vector, and the manifold surface in this direction moves closer to the position with a lower actual response.

[0105] In this embodiment, the manifold parameter refitting unit receives a first correction signal or a second correction signal, and performs an incremental refitting operation on the manifold parameters of the population reference manifold according to the indication of the correction signal. The incremental refitting operation only adjusts the manifold parameters in the population reference manifold that are related to the tangential projection direction of the behavior deviation vector, without recalculating all manifold parameters. The input of the incremental refitting operation is the manifold parameters of the current population reference manifold and the correction signal, and the output is the updated manifold parameters of the population reference manifold.

[0106] In this embodiment, the manifold parameters of the group reference manifold refer to the set of parameters used to define the geometry and spatial position of the group reference manifold. The manifold parameters include the position parameters of each coordinate point in the locally linearly embedded manifold and the connection weight parameters between each coordinate point.

[0107] In this embodiment, the incremental refit operation refers to using the current feedback data as a new constraint to adjust the local parameters of the population reference manifold near the tangential projection direction of the behavior deviation vector. Specifically, in the locally linearly embedded manifold of the population reference manifold, the projected coordinate point and its set of nearest neighbor coordinate points are located. The numerical direction of the correction signal is used as the constraint direction, and a small offset along the constraint direction is applied to the position parameters of the nearest neighbor coordinate points. The magnitude of the offset is proportional to the absolute value of the correction signal, and the direction of the offset is consistent with the positive or negative sign of the correction signal.

[0108] In another embodiment of the invention, it further includes: An edge computing node cluster consists of multiple edge computing nodes distributed in different geographical locations, with each edge computing node communicating with user terminal devices within its corresponding geographical area. The deviation vector local cache unit is set inside each edge computing node to cache the behavior deviation vector and deviation geometric magnitude of each user in the corresponding geographical area of ​​the edge computing node; The local weight pre-correction unit is set inside each edge computing node. When the communication connection between the edge computing node and the group baseline manifold construction module is interrupted, it performs weight correction vector generation and weight correction execution operations locally on the edge computing node based on the locally cached behavior deviation vector and deviation geometric magnitude to generate a local emergency weight configuration. The local emergency weight configuration is replaced by the latest corrected advertising delivery weight configuration generated by the weight correction execution module after the communication connection is restored.

[0109] In this embodiment, the corresponding geographical region refers to the geographical area covered by each edge computing node. The corresponding geographical region is determined according to the physical location and service range of the network deployment, and different edge computing nodes correspond to different geographical regions.

[0110] In this embodiment, the user terminal device refers to the electronic device used by the target user to receive advertising content. The user terminal device includes mobile terminal devices and fixed terminal devices. Mobile terminal devices include smartphones and tablets, and fixed terminal devices include personal computers and smart TVs.

[0111] In this embodiment, the edge computing node group consists of multiple edge computing nodes distributed in different geographical locations. Each edge computing node is deployed in a geographical area. Each edge computing node establishes a communication connection with user terminal devices in the corresponding geographical area through a wireless network or wired network. The edge computing node receives user behavior data reported by the user terminal devices and sends advertising content to the user terminal devices.

[0112] In this embodiment, the deviation vector local cache unit is installed inside each edge computing node. The deviation vector local cache unit obtains the behavior deviation vector and deviation geometric magnitude of each user in the geographical area corresponding to the edge computing node from the behavior deviation vector calculation module, and stores the behavior deviation vector and deviation geometric magnitude in the local storage medium of the edge computing node. The behavior deviation vector and deviation geometric magnitude are indexed in the local storage medium according to the user identifier.

[0113] In this embodiment, communication connection interruption refers to the network link between the edge computing node and the group baseline manifold building module being unavailable. Communication connection interruption includes three situations: network disconnection, server unresponsiveness, and timeout without reply.

[0114] In this embodiment, after detecting an interruption in the communication connection between the edge computing node and the population baseline manifold construction module, the local weight pre-correction unit reads the locally cached behavior deviation vector and deviation geometric magnitude from the deviation vector local cache unit. Based on the behavior deviation vector and deviation geometric magnitude, it performs a weight correction vector generation operation and a weight correction execution operation locally on the edge computing node, generating a corrected advertising delivery weight configuration as a local emergency weight configuration. During the communication connection interruption, the local emergency weight configuration replaces the corrected advertising delivery weight configuration to perform advertising delivery operations. When the communication connection is restored, the local weight pre-correction unit stops using the local emergency weight configuration and replaces it with the latest corrected advertising delivery weight configuration generated by the weight correction execution module, restoring the normal weight configuration execution path.

[0115] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A system for dynamically adjusting ad placement weights based on real-time user behavior characteristics, characterized in that, include: The Behavior State Space Construction Module is used to construct a user behavior state space with behavior type as the dimension axis and behavior time coordinates and intensity coordinates as coordinate components. The user behavior state is expressed as the behavior state coordinate points in the user behavior state space. The user group segmentation module is used to divide all users into multiple user segmentation grid units according to preset segmentation attributes and segmentation levels; The group baseline manifold construction module is used to fit and generate a group baseline manifold based on the set of historical behavioral state coordinates of all users within the hierarchical grid cell of the target user. The behavior deviation vector calculation module is used to determine the target reference point on the group's baseline manifold, and construct the behavior deviation vector with the target reference point as the starting point and the real-time behavior state coordinate point of the target user as the ending point. The weight correction vector generation module is used to decompose the behavior deviation vector into the normal component of the normal direction of the group reference manifold and the tangential component of the tangential direction of the group reference manifold. The direction of the normal component is used as the weight correction direction, and the magnitude of the normal component is used as the basis for calculating the weight correction magnitude to generate the weight correction vector. The weight correction execution module is used to correct the initial ad delivery weight configuration based on the weight correction vector and generate the corrected ad delivery weight configuration. The ad delivery execution module is used to execute ad delivery operations based on the revised ad delivery weight configuration.

2. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 1, characterized in that, The behavior state space construction module includes: The behavior type dimension axis definition unit is used to define each behavior type in the preset behavior type set as a dimension axis of the user behavior state space. The preset behavior type set includes click behavior type, browsing duration behavior type, interaction frequency behavior type, and conversion completion behavior type. The time coordinate component mapping unit is used to normalize the occurrence timestamp of the target behavior according to the preset time window length, and use the normalized time value as the first coordinate component of the target behavior on the corresponding behavior type dimension axis. The intensity coordinate component mapping unit is used to convert the original interaction data of the target behavior into the behavior intensity quantization value according to the preset behavior intensity quantization rule, and use the behavior intensity quantization value as the second coordinate component of the target behavior on the corresponding behavior type dimension axis. The coordinate point generation unit is used to combine the first and second coordinate components on the dimension axis of all behavior types of the target user at the target time to generate the behavior state coordinate points of the target user at the target time.

3. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 1, characterized in that, The group baseline manifold building module includes: The same-layer historical coordinate aggregation unit is used to extract the set of historical behavior state coordinates of all users in the target user's hierarchical grid unit within a preset historical time period from the user database. The manifold structure definition unit is used to define the manifold structure of the group's baseline manifold as a locally linearly embedded manifold, where the embedding dimension of the locally linearly embedded manifold is smaller than the dimension of the user behavior state space. The manifold parameter fitting unit is used to fit the manifold parameters of the local linearly embedded manifold using the set of historical behavior state coordinate points as training data and minimizing the sum of squared projection residuals from each coordinate point in the set of historical behavior state coordinate points to the local linearly embedded manifold as the optimization objective, thereby generating the fitted population baseline manifold.

4. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 1, characterized in that, The behavior deviation vector calculation module includes: The manifold projection calculation unit is used to calculate the projection coordinates of the real-time behavior state coordinates on the group reference manifold based on the local linear reconstruction relationship of the group reference manifold. The projection coordinates are the coordinates on the group reference manifold that are closest to the real-time behavior state coordinates. The historical trajectory positioning unit is used to project each historical behavior state coordinate point onto the group reference manifold surface according to the historical behavior state coordinate point sequence of the target user to obtain a set of historical trajectory projection points on the group reference manifold. The historical projection point in the set of historical trajectory projection points that is closest to the current time is used as the target reference point on the group reference manifold. The deviation vector determination unit is used to construct a behavior deviation vector starting from the target reference point and ending at the real-time behavior state coordinate point.

5. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 4, characterized in that, The weight correction vector generation module includes: The normal direction extraction unit is used to extract the normal direction of the real-time behavior state coordinate point at the projected coordinate point on the group reference manifold, and use the normal direction as the reference direction for weight correction. The normal component decomposition unit is used to project the behavior deviation vector onto the weight correction reference direction to obtain the normal component of the behavior deviation vector in the direction of the group reference manifold normal. The normal magnitude normalization unit is used to compare the magnitude of the normal component with a preset magnitude threshold. When the magnitude of the normal component is less than or equal to the preset magnitude threshold, the weight correction magnitude is set to zero. When the magnitude of the normal component is greater than the preset magnitude threshold, the difference between the magnitude of the normal component and the preset magnitude threshold is used as the weight correction magnitude. The weight correction vector synthesis unit is used to synthesize a weight correction vector by taking the unit direction of the normal component as the weight correction direction component and the weight correction magnitude component as the weight correction magnitude component.

6. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 5, characterized in that, Normal component decomposition elements include: The tangential component calculation sub-unit is used to subtract the normal component from the behavior deviation vector to obtain the tangential component of the behavior deviation vector in the tangential direction of the population reference manifold; The weight correction direction selection subunit is used to compare the magnitudes of the normal and tangential components. When the magnitude of the normal component is greater than that of the tangential component, the direction of the normal component is output as the weight correction direction. When the magnitude of the normal component is less than or equal to that of the tangential component, it is determined that the real-time behavior state of the target user is within the range of the group reference manifold, and the amplitude is set to zero signal to the normal magnitude normalization unit.

7. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 1, characterized in that, The weight adjustment execution module includes: The correction direction mapping unit is used to determine the projection component value of the weight correction vector direction on each behavior type dimension axis based on the direction cosine of the weight correction vector in the user behavior state space, and to determine the ad type corresponding to the behavior type dimension axis with the largest projection component value as the ad type to be corrected according to the preset mapping relationship between the behavior type dimension axis and the ad type. The correction calculation unit is used to multiply the weight correction magnitude by a preset magnitude mapping coefficient to generate the correction value for each type of advertisement to be corrected. The preset magnitude mapping coefficient is positively correlated with the placement intensity parameter in the advertiser's placement target parameters. The weight configuration correction unit is used to add the corresponding correction value to the weight components corresponding to each type of advertisement to be corrected in the initial advertisement placement weight configuration, and generate the corrected advertisement placement weight configuration.

8. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 1, characterized in that, Also includes: The actual response feedback collection module is used to collect the actual delivery response data of target users under the corrected ad delivery weight configuration. The actual delivery response data includes the actual click-through rate and the actual conversion rate. The manifold correction module is used to generate a manifold correction signal based on the difference between the actual delivery response data and the predicted manifold behavior response value. The manifold correction signal is then fed back to the population baseline manifold construction module, which uses the manifold correction signal as a constraint to adjust the manifold parameters of the population baseline manifold. The vector correction module generates a vector correction signal based on the difference between the actual delivery response data and the predicted delivery effect response value. The vector correction signal is then fed back to the weight correction vector generation module to update the calculation mapping parameters of the weight correction magnitude. The predicted delivery effect response value is calculated by the preset delivery effect prediction model based on the corrected advertising delivery weight configuration.

9. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 8, characterized in that, The manifold correction module includes: The prediction response calculation unit is used to calculate the predicted response value of the manifold behavior in the reference manifold by taking the projected coordinates of the real-time behavior state coordinates as the reference, offsetting the behavior deviation vector along the projection direction on the manifold tangent plane and the geodesic of the manifold by a preset step size, and then inputting the weight correction vector corresponding to the offset coordinates into the delivery effect prediction model to obtain the predicted response value of the manifold behavior corresponding to the offset coordinates. The correction signal generation unit is used to calculate the difference between the actual delivery response data and the predicted response value of the manifold behavior. When the difference is positive, a first correction signal is generated, which indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is increased. When the difference is negative, a second correction signal is generated, which indicates that the local curvature of the group reference manifold in the tangential projection direction of the behavior deviation vector is decreased. The manifold parameter refitting unit is used to perform incremental refitting operations on the manifold parameters of the population reference manifold based on a first correction signal or a second correction signal.

10. The system for dynamically adjusting ad placement weights based on real-time user behavior characteristics according to claim 1, characterized in that, Also includes: An edge computing node cluster consists of multiple edge computing nodes distributed in different geographical locations, with each edge computing node communicating with user terminal devices within its corresponding geographical area. The deviation vector local cache unit is set inside each edge computing node to cache the behavior deviation vector and deviation geometric magnitude of each user in the corresponding geographical area of ​​the edge computing node; The local weight pre-correction unit is set inside each edge computing node. When the communication connection between the edge computing node and the group baseline manifold construction module is interrupted, it performs weight correction vector generation and weight correction execution operations locally on the edge computing node based on the locally cached behavior deviation vector and deviation geometric magnitude to generate a local emergency weight configuration. The local emergency weight configuration is replaced by the latest corrected advertising delivery weight configuration generated by the weight correction execution module after the communication connection is restored.

Images