Cross-platform user behavior analysis and media information matching method and system

Abstract

The application provides a cross-platform user behavior analysis and media information matching method and system, relates to the technical field of user analysis, and comprises the following steps: time sequence analysis of multi-platform user behavior data, generation of a dynamic interest vector with time decay weight, and mapping to a unified semantic space to obtain user potential interest distribution. Based on this, a multi-granularity user group portrait is constructed, and a sequential decision model with long-term conversion revenue as the target is used to dynamically generate platform selection, delivery timing and content matching strategies. After the delivery is executed, the model parameters and decay rules are updated using feedback data to form an optimized closed loop. The application realizes accurate capture of dynamic evolution of user interest and adaptive optimization of cross-platform delivery strategy, and improves marketing efficiency and return on investment.

Description

Technical Field

[0001] This invention relates to user analytics technology, and more particularly to a method and system for cross-platform user behavior analysis and media information matching. Background Technology

[0002] In the fields of digital marketing and cross-platform advertising, existing technologies typically rely on the collection and analysis of user behavior data. The conventional approach involves obtaining user behavior logs such as clicks, browsing, and purchases from multiple independent platforms and attempting to build a unified user profile. These methods often involve simply aggregating or weighting data from different sources to form a static set of user interest tags. For ad delivery, the system matches these static tags with the keywords or categories of the ad content and, within preset budget and time constraints, completes ad display through a real-time bidding mechanism.

[0003] Existing technologies have significant shortcomings. On the one hand, conventional methods often process user behavior data in a static and fragmented manner. They fail to fully consider the dynamic evolution of user interests over time, such as the phased shifts in interests, the decay of trending topics, and the inherent correlations between behavioral patterns across different platforms. Simply aggregating data from different platforms ignores the semantic gaps and temporal asynchronicity between heterogeneous data sources, resulting in rigid and outdated user profiles that cannot accurately reflect the real-time, cross-platform comprehensive interest status of users. On the other hand, most existing ad placement decision models focus on maximizing the immediate benefits of a single impression or simply allocating budgets under fixed rules. This strategy lacks a long-term, global optimization perspective and struggles to achieve Pareto optimality in long-term conversion benefits and budget efficiency in complex multi-platform bidding environments, fluctuating traffic, and dynamic scenarios where user interests are constantly changing. Furthermore, an effective closed loop is not formed between the decision-making process and actual user feedback, leading to insufficient adaptive capabilities of the model strategy and a tendency for ad placement performance to deteriorate over time. Summary of the Invention

[0004] This invention provides a method and system for cross-platform user behavior analysis and media information matching, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a cross-platform user behavior analysis and media information matching method, comprising:

[0006] User behavior data from multiple heterogeneous platforms is acquired, and time-series analysis is performed on the user behavior data. By capturing the evolution of user behavior patterns under different time windows, the phased transfer characteristics and decay patterns of user interests are identified, a dynamic interest representation vector with time decay weight is generated, and the heterogeneous behavior features of different platforms are mapped to a unified semantic space to obtain the distribution of user potential interests that integrates information from multiple platforms.

[0007] Based on the matching relationship between the user's potential interest distribution and the semantic features of the platform content, a multi-granularity user group profile is constructed.

[0008] Establish a campaign decision model. Based on the campaign decision model, take the multi-granular user group profile, platform bidding environment, and time period traffic distribution as state inputs. Through a sequential decision optimization framework, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints, dynamically generate platform selection strategies, campaign timing allocation schemes, and creative content matching rules for different user groups. And continuously evolve the decision strategy based on the cumulative feedback of historical campaign performance.

[0009] Based on the output strategy of the aforementioned delivery decision model, targeted delivery is executed on the corresponding platform, and multi-dimensional feedback data is collected. The time decay weight calculation rules in the time series analysis process are updated using the multi-dimensional feedback data, and the strategy generation parameters of the delivery decision model are adjusted at the same time, forming a two-way closed loop of user understanding and delivery optimization.

[0010] Time-series analysis is performed on the user behavior data. By capturing the evolution of user behavior patterns in different time windows, the phased shift characteristics and decay patterns of user interests are identified, and a dynamic interest representation vector with time decay weights is generated, including:

[0011] User behavior data is divided into multiple consecutive time windows according to the time dimension, and behavioral feature vectors are extracted for the user interaction sequence within each time window.

[0012] Establish a behavioral evolution correlation map between time windows, identify the migration trajectory of user interests between different topic domains by calculating the semantic distance and distribution offset of behavioral feature vectors between adjacent time windows, and determine the stage boundary of interest transfer based on the persistence and retrospectiveness of the migration trajectory, so as to obtain the stage transfer feature identifier that divides the interest evolution stage.

[0013] For each interest evolution stage in the phased transition feature identifier, a two-factor decay function is constructed based on the time span between the interest evolution stage and the current time and the fluctuation amplitude of the behavioral feature vector within the interest evolution stage. The time span is mapped to a time distance decay factor and the behavioral fluctuation amplitude is mapped to a stability decay factor according to the two-factor decay function. The final decay coefficient is calculated by fusing the time distance decay factor and the stability decay factor to generate a time decay weight sequence that matches the characteristics of the interest evolution stage.

[0014] The behavioral feature vector is fused with the time decay weight sequence at corresponding positions to generate a dynamic interest representation vector with time decay weights.

[0015] Based on the matching relationship between the user's potential interest distribution and the semantic features of the platform content, a multi-granularity user group profile is constructed, including:

[0016] The distribution of users’ potential interests is mapped to the semantic feature space of platform content. By constructing an interest-content matching metric function, the response strength between each interest dimension in the user’s interest distribution and the semantic features of platform content is calculated. The response strength reflects both the sensitivity of the interest dimension to a specific content type and the triggering ability of the content feature to activate the interest, thus obtaining an interest-content matching relationship matrix that characterizes the individual content consumption preferences of users.

[0017] Based on the interest-content matching relationship matrix, cross-user response pattern analysis is performed. By identifying the distribution pattern of response intensity and activation path similarity of different users in the interest-content matching relationship matrix, users with similar response patterns are aggregated into user groups. For each user group, common response patterns and specific response patterns are extracted from the matching relationship matrix of its members to generate group-level feature identifiers that distinguish different groups.

[0018] By analyzing the overlapping areas and evolution trends of response patterns among the group-level feature identifiers, and by tracking the change trajectory of user response patterns in different groups, a cross-group association map describing the migration patterns of users between groups is constructed.

[0019] The interest-content matching matrix, the group-level feature identifier, and the cross-group association graph are combined to form a multi-granularity user group profile.

[0020] Analyzing the overlapping areas and evolution trends of response patterns among the group-level feature identifiers, and by tracking the changing trajectories of user response patterns across different groups, a cross-group association map describing the migration patterns of users between groups is constructed, including:

[0021] Cross-group response pattern spatial projection is performed on group-level feature identifiers. By constructing a unified response pattern representation space and mapping the response patterns of different groups to the unified response pattern representation space, the spatial overlap region and separation boundary of response patterns between groups are calculated. Group pairs with continuous response features and group pairs with broken response are identified, and the topological structure of the overlapping region that quantifies the similarity and difference of responses between groups is obtained.

[0022] Based on the overlapping region topology, the user's group affiliation evolution trajectory in the time series is tracked. By analyzing the temporal movement path of the user response pattern in the unified response pattern representation space, the migration trajectory of the user from the source group response region to the target group response region is identified. Based on the spatial span, migration speed and trajectory density distribution of the migration trajectory, the directional features of user migration between groups, migration concentration and response pattern change threshold are extracted.

[0023] Using the population pairs in the overlapping region topology as graph nodes, the directional features of the migration trajectory and the migration concentration as directed connection edge attributes between nodes, and the response pattern change threshold as the edge activation condition, a cross-population association graph is constructed.

[0024] By employing a sequential decision optimization framework aimed at maximizing long-term conversion benefits and achieving Pareto optimality within budget constraints, the framework dynamically generates platform selection strategies, ad placement timing allocation schemes, and creative content matching rules for different user groups, including:

[0025] A sequential decision optimization framework is established, and the campaign process is divided into multiple consecutive decision stages according to the sequential decision optimization framework. In each decision stage, a state-action-value mapping relationship is constructed to predict the fluctuation of the platform bidding environment and the migration of traffic distribution triggered by the campaign action targeting a specific user group in the current stage. By retrospectively tracing the cumulative impact of the conversion revenue contribution and budget consumption burden of future stages on the value of the action in the current stage, the long-term conversion revenue expectation and budget consumption cumulative trajectory of the campaign action in the current stage are quantified.

[0026] The expected long-term conversion revenue is mapped to a dual-objective space along with the cumulative budget consumption trajectory, and a dynamic Pareto front is constructed. By iteratively adjusting the sequence of delivery actions across decision-making stages, the projection point of the action in the dual-objective space gradually approaches the dynamic Pareto front. On the dynamic Pareto front, a non-dominated solution that satisfies the budget constraint boundary conditions and maximizes the expected long-term conversion revenue is selected. The delivery action sequence corresponding to the non-dominated solution is then analyzed into platform selection strategies, delivery timing allocation schemes, and creative content matching rules for different user groups at each decision-making stage.

[0027] Mapping the expected long-term conversion revenue and the cumulative trajectory of budget expenditure to a dual-objective space and constructing a dynamic Pareto front includes:

[0028] A dual-objective space is established, with the expected value of long-term conversion revenue as the conversion revenue dimension and the cumulative trajectory of budget consumption as the budget consumption dimension. The expected value of long-term conversion revenue and the cumulative trajectory of budget consumption corresponding to each candidate delivery action sequence in the sequential decision optimization framework are used as coordinate components, and the coordinate components are mapped to the dual-objective space to form a distribution point set.

[0029] For each candidate solution in the distribution point set, count the number of other candidate solutions that are superior to the candidate solution in both the transformation benefit dimension and the budget consumption dimension, quantify their degree of domination, and select candidate solutions with a degree of domination of zero to form a non-dominated solution set. Define the boundary curve formed by the non-dominated solution set in the dual objective space as the dynamic Pareto front.

[0030] Based on the output strategy of the aforementioned delivery decision model, targeted delivery is executed on the corresponding platform, and multi-dimensional feedback data is collected. The time decay weight calculation rules in the time series analysis process are updated using the multi-dimensional feedback data, including:

[0031] Based on the output strategy of the delivery decision model, targeted delivery is executed on the corresponding platform. During the delivery process, multi-dimensional feedback data is collected, and the time delay distribution characteristics of user response and the time-series fluctuation characteristics of platform bidding are extracted from the multi-dimensional feedback data.

[0032] Based on the time delay distribution characteristics and the time series fluctuation characteristics, an attenuation relationship is constructed, the attenuation relationship is transformed into a weight calculation rule, and the time attenuation weight calculation rule in the time series analysis process is updated using the weight calculation rule.

[0033] A second aspect of the present invention provides a cross-platform user behavior analysis and media information matching system, comprising:

[0034] The data fusion unit is used to acquire user behavior data from multiple heterogeneous platforms, perform time-series analysis on the user behavior data, capture the evolution of user behavior patterns under different time windows, identify the phased transfer characteristics and decay patterns of user interests, generate dynamic interest representation vectors with time decay weights, and map the heterogeneous behavior features of different platforms to a unified semantic space to obtain the distribution of potential user interests that integrates information from multiple platforms.

[0035] The profile building unit is used to build multi-granularity user group profiles based on the matching relationship between the distribution of potential user interests and the semantic features of platform content;

[0036] The decision optimization unit is used to establish a campaign decision model. Based on the campaign decision model, the multi-granular user group profile, platform bidding environment, and time period traffic distribution are used as state inputs. Through the sequential decision optimization framework, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints, the unit dynamically generates platform selection strategies, campaign timing allocation schemes, and creative content matching rules for different user groups. The decision strategy is continuously evolved based on the cumulative feedback of historical campaign performance.

[0037] The closed-loop execution unit is used to execute targeted delivery on the corresponding platform according to the output strategy of the delivery decision model, collect multi-dimensional feedback data, update the time decay weight calculation rules in the time series analysis process using the multi-dimensional feedback data, and adjust the strategy generation parameters of the delivery decision model to form a two-way closed loop of user understanding and delivery optimization.

[0038] A third aspect of the present invention provides an electronic device, comprising:

[0039] processor;

[0040] Memory used to store processor-executable instructions;

[0041] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0043] The beneficial effects of this application are as follows:

[0044] This method enables a deep understanding and dynamic modeling of user behavior across multiple platforms. By capturing the phased shifts and decay patterns of user interests through time-series analysis, it generates dynamic interest representation vectors with time decay weights. This effectively overcomes the shortcomings of traditional static profiles in reflecting user interest drift, significantly improving the timeliness and accuracy of user interest representation. By mapping heterogeneous behavioral features from different platforms to a unified semantic space, it breaks down data barriers between platforms, achieving deep fusion of multi-source information and thus obtaining a more comprehensive and accurate distribution of potential user interests.

[0045] By constructing multi-granular user profiles based on a unified distribution of potential user interests, user segmentation becomes more refined and comprehensive. This method considers not only macro-level group commonalities but also micro-level individual differences, providing a solid data foundation for subsequent precise targeting. The constructed profiles can adaptively reflect the dynamic changes of user groups, ensuring a consistently high relevance between marketing strategies and the target audience.

[0046] This method establishes a sequential decision optimization framework, enabling intelligent and dynamic ad placement strategies. By integrating multi-granular user profiles, real-time bidding environment, and time-based traffic distribution as comprehensive inputs, it makes decisions with the goal of maximizing long-term conversion returns, avoiding local optima caused by optimizing a single campaign and ensuring the long-term return on the overall marketing budget. Under budget constraints, the model can automatically seek Pareto optimal solutions and dynamically generate joint strategies for platform selection, timing of placement, and content matching, significantly improving the efficiency and effectiveness of advertising resource allocation. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the cross-platform user behavior analysis and media information matching method according to an embodiment of the present invention;

[0048] Figure 2 This is a flowchart illustrating the cross-platform user interest modeling and media information matching decision-making process in an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0051] Figure 1 This is a flowchart illustrating the cross-platform user behavior analysis and media information matching method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0052] User behavior data from multiple heterogeneous platforms is acquired, and time-series analysis is performed on the user behavior data. By capturing the evolution of user behavior patterns under different time windows, the phased transfer characteristics and decay patterns of user interests are identified, a dynamic interest representation vector with time decay weight is generated, and the heterogeneous behavior features of different platforms are mapped to a unified semantic space to obtain the distribution of user potential interests that integrates information from multiple platforms.

[0053] Based on the matching relationship between the user's potential interest distribution and the semantic features of the platform content, a multi-granularity user group profile is constructed.

[0054] Establish a campaign decision model. Based on the campaign decision model, take the multi-granular user group profile, platform bidding environment, and time period traffic distribution as state inputs. Through a sequential decision optimization framework, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints, dynamically generate platform selection strategies, campaign timing allocation schemes, and creative content matching rules for different user groups. And continuously evolve the decision strategy based on the cumulative feedback of historical campaign performance.

[0055] Based on the output strategy of the aforementioned delivery decision model, targeted delivery is executed on the corresponding platform, and multi-dimensional feedback data is collected. The time decay weight calculation rules in the time series analysis process are updated using the multi-dimensional feedback data, and the strategy generation parameters of the delivery decision model are adjusted at the same time, forming a two-way closed loop of user understanding and delivery optimization.

[0056] In one optional implementation, time-series analysis is performed on the user behavior data. By capturing the evolution of user behavior patterns under different time windows, the phased shift characteristics and decay patterns of user interests are identified, and a dynamic interest representation vector with time decay weights is generated, including:

[0057] User behavior data is divided into multiple consecutive time windows according to the time dimension, and behavioral feature vectors are extracted for the user interaction sequence within each time window.

[0058] Establish a behavioral evolution correlation map between time windows, identify the migration trajectory of user interests between different topic domains by calculating the semantic distance and distribution offset of behavioral feature vectors between adjacent time windows, and determine the stage boundary of interest transfer based on the persistence and retrospectiveness of the migration trajectory, so as to obtain the stage transfer feature identifier that divides the interest evolution stage.

[0059] For each interest evolution stage in the phased transition feature identifier, a two-factor decay function is constructed based on the time span between the interest evolution stage and the current time and the fluctuation amplitude of the behavioral feature vector within the interest evolution stage. The time span is mapped to a time distance decay factor and the behavioral fluctuation amplitude is mapped to a stability decay factor according to the two-factor decay function. The final decay coefficient is calculated by fusing the time distance decay factor and the stability decay factor to generate a time decay weight sequence that matches the characteristics of the interest evolution stage.

[0060] The behavioral feature vector is fused with the time decay weight sequence at corresponding positions to generate a dynamic interest representation vector with time decay weights.

[0061] When performing time-series analysis of user behavior data, the raw data is divided into fixed-duration time windows based on the time of occurrence of user behavior. The length of the time window is determined according to the platform characteristics; for example, it can be set to a 24-hour window for e-commerce platforms and a 2-hour window for short video platforms. For user interaction behaviors recorded within each time window, including operation types such as clicks, browsing, favorites, and sharing, corresponding attribute information such as content category, keyword tags, and dwell time is extracted. This attribute information is encoded into a fixed-dimensional numerical vector as the behavior feature vector for that time window.

[0062] When constructing a behavioral evolution correlation graph, each time window is considered a node in the graph, and edges are established between nodes using a semantic distance metric. The semantic distance is calculated using cosine similarity or Euclidean distance formulas, specifically expressed as follows: , where v t and v t+1These are the behavioral feature vectors for adjacent time windows. When the semantic distance exceeds a preset threshold, it is marked as a significant shift event. Further analysis of whether the shift directions of multiple consecutive windows remain consistent identifies the migration trajectory. If the components of the behavioral feature vectors in a specific topic domain continuously increase for three or more consecutive time windows, it is determined that there is an interest shift towards that topic. The stage boundary is detected based on the rate of change of the shift degree. When the rate of change falls from a high level to a stable range, it is marked as the end point of an interest evolution stage, completing the division of stage-based transfer feature identification.

[0063] When calculating the time decay weight, for each identified interest evolution stage, the time difference Δt between the start timestamp of that stage and the current time is obtained, and the time distance decay factor is calculated. , where α is the decay rate parameter. Simultaneously, the variance of all behavioral feature vectors within this stage is calculated as the fluctuation amplitude index σ. 2 The stability decay factor is obtained through normalization. Where k is the sensitivity adjustment coefficient. The final attenuation coefficient is obtained by fusing the two types of attenuation factors in the form of a weighted product. The exponent γ is used to adjust the strength of the stability factor. A decay coefficient is calculated for each stage of interest evolution, and these coefficients are arranged in chronological order to form a time-decay weighted sequence.

[0064] When generating the dynamic interest representation vector, the behavioral feature vectors of each interest evolution stage are multiplied element-wise with their corresponding decay weights to obtain a weighted set of feature vectors. Pooling is then performed on all weighted feature vectors, aggregating them into a single vector using either weighted summation or max pooling. This vector is the dynamic interest representation vector that incorporates time decay information, maintaining the same dimension as the original behavioral feature vectors and reflecting the differentiated weight distribution of users' recent and long-term interests.

[0065] In one optional implementation, constructing a multi-granularity user group profile based on the matching relationship between the user's potential interest distribution and the semantic features of platform content includes:

[0066] The distribution of users’ potential interests is mapped to the semantic feature space of platform content. By constructing an interest-content matching metric function, the response strength between each interest dimension in the user’s interest distribution and the semantic features of platform content is calculated. The response strength reflects both the sensitivity of the interest dimension to a specific content type and the triggering ability of the content feature to activate the interest, thus obtaining an interest-content matching relationship matrix that characterizes the individual content consumption preferences of users.

[0067] Based on the interest-content matching relationship matrix, cross-user response pattern analysis is performed. By identifying the distribution pattern of response intensity and activation path similarity of different users in the interest-content matching relationship matrix, users with similar response patterns are aggregated into user groups. For each user group, common response patterns and specific response patterns are extracted from the matching relationship matrix of its members to generate group-level feature identifiers that distinguish different groups.

[0068] By analyzing the overlapping areas and evolution trends of response patterns among the group-level feature identifiers, and by tracking the change trajectory of user response patterns in different groups, a cross-group association map describing the migration patterns of users between groups is constructed.

[0069] The interest-content matching matrix, the group-level feature identifier, and the cross-group association graph are combined to form a multi-granularity user group profile.

[0070] After obtaining the distribution of users' potential interests, it is projected onto the platform's content semantic feature space. The platform's content semantic features are obtained by encoding the text, tags, and classification system of historically published content, forming a feature vector set of dimension M. The distribution of users' potential interests is represented as an N-dimensional vector, where each dimension corresponds to a category of interest topics. When constructing the interest-content matching metric function, the semantic similarity between the user's interest vector and the content feature vector is calculated using a cosine distance metric, incorporating a time decay weight for the interest dimension as an adjustment factor. Response intensity is obtained through bidirectional calculation: forward calculation assesses the sensitivity of the user's interest dimension to specific content types, reflecting the user's tendency to actively search for or browse such content; backward calculation assesses the activation ability of content features on user interests, measuring the trigger probability of specific content generating attention within the user's interest spectrum. For a single user, the combinations of each dimension of their interest distribution with all content feature types on the platform are traversed to form an N×M interest-content matching relationship matrix, where the matrix elements represent the response intensity of the corresponding interest dimension and content feature.

[0071] After collecting matching relationship matrices from multiple users, cross-user response pattern analysis is performed. Elements in each user's matrix with response intensities exceeding a preset threshold are extracted, forming the user's activation path set. The intersection ratio and statistical distance of response intensity distributions between different user activation path sets are calculated. A density-based clustering algorithm is used to group users with similar response patterns into the same group. For each user group, the mean and variance of response intensity at each position in the matching relationship matrix of its members are statistically analyzed. Positions with high mean and low variance reflect common response patterns within the group, representing the group's general content consumption preferences; positions with medium mean but high variance reflect specific response patterns, indicating differentiation characteristics within the group. The interest-content combinations corresponding to the common response patterns of the groups are encoded into keyword sequences and combined with the distribution feature parameters of the specific response patterns of the groups to generate group-level feature identifiers that distinguish different groups.

[0072] This study analyzes the distribution of group-level feature identifiers in the interest-content space and identifies overlapping regions of different group feature identifiers in the matrix space. The response patterns of overlapping regions represent the common interests of different groups. The study calculates the trend of response intensity changes in overlapping regions to determine the convergence or divergence of interests between groups. It tracks the changes in the matching relationship matrix of individual users within a continuous time window, detecting the trajectory of their response patterns migrating from one group feature to another. The study statistically analyzes the frequency of user migration from group A to group B across time periods, as well as the transitional characteristics of response patterns during the migration process. A cross-group association graph is established with groups as nodes and migration frequency as edge weights. The edge directions in the graph indicate the main flow of user migration.

[0073] The interest-content matching matrix serves as the individual-level profile, group-level feature identifiers serve as the group-level profile, and cross-group association graphs serve as the ecosystem-level profile, combining to form a multi-granularity user group profile. This profile allows for on-demand selection of analysis granularity: the individual-level matrix is ​​used for precise targeting scenarios, group-level feature identifiers are used for large-scale targeting, and the cross-group association graph is referenced for long-term strategy planning.

[0074] In one optional implementation, analyzing the overlapping areas and evolution trends of response patterns among the group-level feature identifiers, and constructing a cross-group association map describing the migration patterns of users between different groups by tracking the changing trajectories of user response patterns across different groups includes:

[0075] Cross-group response pattern spatial projection is performed on group-level feature identifiers. By constructing a unified response pattern representation space and mapping the response patterns of different groups to the unified response pattern representation space, the spatial overlap region and separation boundary of response patterns between groups are calculated. Group pairs with continuous response features and group pairs with broken response are identified, and the topological structure of the overlapping region that quantifies the similarity and difference of responses between groups is obtained.

[0076] Based on the overlapping region topology, the user's group affiliation evolution trajectory in the time series is tracked. By analyzing the temporal movement path of the user response pattern in the unified response pattern representation space, the migration trajectory of the user from the source group response region to the target group response region is identified. Based on the spatial span, migration speed and trajectory density distribution of the migration trajectory, the directional features of user migration between groups, migration concentration and response pattern change threshold are extracted.

[0077] Using the population pairs in the overlapping region topology as graph nodes, the directional features of the migration trajectory and the migration concentration as directed connection edge attributes between nodes, and the response pattern change threshold as the edge activation condition, a cross-population association graph is constructed.

[0078] When projecting group-level feature identifiers across group response pattern spaces, a unified response pattern representation space is constructed as the common projection basis for all group response patterns. The unified response pattern representation space adopts a multi-dimensional Euclidean space structure, with the number of spatial dimensions consistent with the number of semantic feature categories on the platform, typically ranging from 20 to 100 dimensions. Each dimension in the space corresponds to an independent content semantic feature category, and the coordinate axes are normalized to between 0 and 1. The set of feature position indices and weight coefficient key-value pairs contained in the group-level feature identifiers is mapped to the unified response pattern representation space in the following way: the feature position index is parsed into a tuple of interest dimension number and content item number; the representation intensity vector of the content item on each semantic feature category is queried; this vector is multiplied by the weight coefficient corresponding to the feature position to obtain the contribution vector of that feature position in the unified space. The contribution vectors of all feature positions in the group-level feature identifiers are summed dimension-wise to obtain the group response vector of that group in the unified response pattern representation space. The vector length is equal to the number of spatial dimensions, and each component represents the comprehensive response intensity of the group on the corresponding semantic feature category.

[0079] When calculating the spatial overlap region and separation boundary of response patterns between groups, vector similarity calculation and difference vector analysis are performed on the group response vectors of any two groups. Cosine similarity is used to measure vector similarity, calculated by dividing the inner product of the two group response vectors by the product of their magnitudes. The similarity value ranges from -1 to 1; the closer the value is to 1, the more similar the overall response patterns of the two groups are. The difference vector is obtained by subtracting the two group response vectors dimension by dimension. The absolute value of each component in the difference vector represents the difference in response intensity between the two groups in that semantic feature dimension. The spatial overlap region consists of the set of dimensions where the values ​​in both group response vectors are above a threshold and the absolute values ​​of the corresponding components in the difference vector are below the threshold. The high response threshold is set by default to 0.8 times the mean of the response vector components of each group, and the difference threshold is set by default to 0.5 times the standard deviation of the difference vector components. The separation boundary consists of the set of dimensions where the absolute value of the difference vector exceeds 1.5 times the difference threshold; the response intensity of the two groups shows significant differences in these dimensions.

[0080] When identifying groups exhibiting continuous response characteristics and those exhibiting broken response characteristics, the continuity type of the group pair is determined by the ratio of the number of dimensions in the spatially overlapping region to the total number of dimensions. A ratio exceeding 0.3 indicates continuous response characteristics, while a ratio below 0.15 indicates broken response characteristics. For continuous groups, the average response intensity difference across the overlapping region dimensions is further calculated as a continuity strength index; the smaller the difference, the stronger the continuity. For broken groups, the number of separation boundary dimensions and the average response intensity difference across those dimensions are used as a breakage degree index. The resulting overlapping region topology is stored as a graph structure, with nodes corresponding to group identifiers. Edge attributes include the overlapping region dimension set, the separation boundary dimension set, the continuity strength index or breakage degree index, and a cosine similarity value. Edges in the topology are tagged as either continuous or broken edges; continuous edges correspond to continuous groups, and broken edges correspond to broken groups.

[0081] When tracking the evolution of user group affiliation over time based on overlapping region topology, the time axis is divided into fixed-length time windows, with the window length set to 7, 14, or 30 days depending on the business cycle. For each user's interest-content matching matrix within each time window, group affiliation is determined. The determination method involves calculating the similarity between the user matrix and all group-level feature identifiers. This similarity is obtained by flattening the user matrix into a response vector and calculating the cosine similarity with the group response vectors. The group with the highest similarity is determined as the user's affiliation group for that time window. The affiliation group identifiers of the same user in adjacent time windows are connected chronologically to form the user's group affiliation sequence. When the affiliation group identifier changes between adjacent time windows, a group migration event is recorded. The event attributes include the source group identifier, target group identifier, migration occurrence time window number, and user identifier.

[0082] When analyzing the temporal movement path of user response patterns in the unified response pattern representation space, for users undergoing group migration, user response vectors for three consecutive time windows before and after migration are extracted. User response vectors are obtained by mapping the matching relationship matrix of the user within that time window to the unified response pattern representation space, using the same mapping method as the generation of group response vectors. The user response vectors from each time window before and after migration are arranged chronologically to form a set of temporal trajectory points for the user in the unified space. The Euclidean distance between adjacent trajectory points is calculated as the single-step movement distance, which is equal to the square root of the sum of the squares of the dimension-wise differences between the user response vectors of two adjacent time windows. When identifying the user's migration trajectory from the source group response region to the target group response region, the spatial region around the source group response vector with an Euclidean distance less than 0.3 times the magnitude of the source group response vector is defined as the source group response region, and the spatial region around the target group response vector with an Euclidean distance less than 0.3 times the magnitude of the target group response vector is defined as the target group response region. The trajectory segment from the user's location in the source group response region to their first entry into the target group response region is extracted as the complete migration trajectory.

[0083] When extracting directional features, migration concentration, and response pattern change thresholds for user migration between groups based on the spatial span, migration speed, and trajectory density distribution of migration trajectories, the spatial span is defined as the Euclidean distance between the starting and ending points of the migration trajectory, reflecting the overall magnitude of change in user response patterns during migration. Migration speed is defined as the spatial span divided by the number of time windows traversed by the migration trajectory, reflecting the rate of change in user response patterns over time. Trajectory density distribution is obtained by dividing the unified response pattern representation space into grid cells and counting the number of migration trajectories traversed within each grid cell. The side length of the grid cell is set to 0.05 times the range of spatial coordinate axes. Directional features are obtained by calculating the average displacement vector of the endpoint of all user trajectories migrating from the source group to the target group relative to the starting point; this vector represents the dominant direction of user migration between the groups. Migration concentration is obtained by calculating the spatial dispersion of all migration trajectory endpoints around the average endpoint position, using the standard deviation of the endpoint coordinates as a measure of dispersion; the smaller the standard deviation, the higher the migration concentration. The response pattern change threshold is defined as the median of the spatial span of all migration trajectories. A significant change in response pattern is determined when the distance the user's response vector moves in adjacent time windows exceeds this threshold.

[0084] When using groups in the overlapping region topology as graph nodes, node attributes include not only the group identifier but also the values ​​of each dimension of the group response vector, the group size (i.e., the number of users belonging to the group), and the group activity level (i.e., the average content consumption frequency of users within the group). When using the directional features and migration concentration of migration trajectories as attributes of directed connections between nodes, the edges point from the source group node to the target group node. The edge attribute fields include the values ​​of each dimension of the directional feature vector, the migration concentration value, the number of migrating users, the average migration speed, and the mean and standard deviation of the spatial span. When using the response pattern change threshold as the edge activation condition, an activation threshold field is set for each edge, with the field value equal to the median of the spatial span of all migration trajectories between the group pairs. When the response vector of a new user moves a distance exceeding this activation threshold, the edge is activated, triggering the user migration prediction mechanism.

[0085] The constructed cross-group association graph is stored using an adjacency list structure. Each node maintains an outgoing edge list and an incoming edge list. The outgoing edge list stores the set of edges and their attributes pointing from that node to other nodes, while the incoming edge list stores the set of edges and their attributes pointing from other nodes to that node. The graph supports operations such as querying node attributes by node identifier, querying edge attributes by source and target nodes, and filtering edges by activation threshold. Query operations are implemented through a hash mapping from node identifier to node storage address, resulting in constant time complexity. Graph update operations are performed after each time window, updating the number of migrating users on edges based on newly generated group migration events, recalculating directional features and migration concentration, and updating activation thresholds.

[0086] In one optional implementation, a sequential decision optimization framework is used to dynamically generate platform selection strategies, ad placement timing allocation schemes, and creative content matching rules for different user groups, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints.

[0087] A sequential decision optimization framework is established, and the campaign process is divided into multiple consecutive decision stages according to the sequential decision optimization framework. In each decision stage, a state-action-value mapping relationship is constructed to predict the fluctuation of the platform bidding environment and the migration of traffic distribution triggered by the campaign action targeting a specific user group in the current stage. By retrospectively tracing the cumulative impact of the conversion revenue contribution and budget consumption burden of future stages on the value of the action in the current stage, the long-term conversion revenue expectation and budget consumption cumulative trajectory of the campaign action in the current stage are quantified.

[0088] The expected long-term conversion revenue is mapped to a dual-objective space along with the cumulative budget consumption trajectory, and a dynamic Pareto front is constructed. By iteratively adjusting the sequence of delivery actions across decision-making stages, the projection point of the action in the dual-objective space gradually approaches the dynamic Pareto front. On the dynamic Pareto front, a non-dominated solution that satisfies the budget constraint boundary conditions and maximizes the expected long-term conversion revenue is selected. The delivery action sequence corresponding to the non-dominated solution is then analyzed into platform selection strategies, delivery timing allocation schemes, and creative content matching rules for different user groups at each decision-making stage.

[0089] like Figure 2 As shown, the method includes:

[0090] When establishing a sequential decision-making optimization framework, the campaign process is divided into continuous decision-making stages with a fixed time granularity. The time granularity is set to 1 hour, 6 hours, or 24 hours based on the budget refresh frequency, and the total campaign duration is 7 to 30 days, corresponding to 168 to 720 stages. A predefined user group set includes 5 to 20 groups, a campaign platform set includes 3 to 8 platforms, and a creative content set includes 10 to 50 content units.

[0091] When constructing the state-action-value mapping relationship at each decision-making stage, the state is defined as a fixed-length numerical vector containing environmental variables and historical cumulative indicators. Environmental variables include the real-time bidding floor price of each platform, traffic supply, peak online time period identifiers for users, activity index of each user group, remaining budget amount, percentage of budget already consumed, and current stage number. Historical cumulative indicators include cumulative conversions, cumulative impressions, cumulative clicks, historical conversion rate of each user group, and historical click-through rate of each platform. The state vector length is 30 to 80 dimensions, continuous variables are normalized to the 0-1 range, and discrete variables are transformed through one-hot encoding.

[0092] The action is defined as the resource allocation plan for each user group in the current phase, including platform selection decisions, timing allocation decisions, and creative content matching decisions. The platform selection decision assigns one or more platforms to each user group, using a hash table to store the mapping from user group identifiers to platform identifier sets. The timing allocation decision divides the current phase into 4 to 12 time slices, generating a time slice weight vector for each user group. The vector length equals the number of time slices, and each component represents the budget allocation percentage for that time slice; the sum of all components equals 1. The creative content matching decision assigns several creative content units and weights to each user group, stored as a mapping from user group identifiers to a creative content weight dictionary. To reduce the action space size, each user group is limited to selecting a maximum of 3 platforms and 5 creative content units. The components of the timing allocation weight vector are selected from a discrete set containing 0, 0.1, 0.2, 0.3, 0.4, and 0.5.

[0093] The value function is defined as the expected long-term conversion return obtained by performing a specific action from the current stage and following a specific strategy in subsequent stages. The expected long-term conversion return equals the sum of the immediate conversion return in the current stage and the discounted conversion returns in all future stages. The immediate conversion return is the number of conversions multiplied by the single conversion return, which ranges from 10 to 500 currency units. The discounted conversion return is the sum of the future stage conversion returns multiplied by a discount factor, with the discount factor ranging from 0.9 to 0.99.

[0094] To predict platform bidding environment fluctuations and traffic distribution migrations triggered by current campaign actions, a platform bidding environment response model and a traffic distribution migration model are constructed. The platform bidding environment response model receives the current state vector and action, and outputs predicted bid prices and traffic supply for each platform in the next stage. The model uses a regression structure, with training data derived from historical campaign logs. The loss function is the mean squared error between predicted and actual values, and the learning rate is set to 0.001 to 0.01, with 100 to 500 training iterations. The traffic distribution migration model receives the current state vector, action, and the output of the platform bidding environment response model, and outputs predicted activity indices for each user group in the next stage and predicted traffic allocation ratios for each platform. It employs a multi-task learning structure, with the loss function being a weighted sum of activity prediction loss and traffic allocation prediction loss, with default weight coefficients of 0.5 and 0.5.

[0095] When tracing the cumulative impact of future stage conversion revenue contribution and budget consumption burden on the value of current stage actions, a dynamic programming algorithm is used to calculate backwards from the final decision stage to the current stage. The value function of the final stage equals the immediate conversion revenue of that stage minus the immediate budget consumption, where the immediate budget consumption is the number of impressions multiplied by the weighted average of the platform's bid floor price. The value function of the penultimate stage equals the immediate conversion revenue of that stage plus the final stage's value function multiplied by a discount factor, minus the immediate budget consumption of that stage. This rule is then used to calculate backwards back to the first stage.

[0096] When quantifying the expected long-term conversion revenue and cumulative budget consumption trajectory of the current stage's campaign actions, the expected long-term conversion revenue is represented by the current stage's value function value, and the cumulative budget consumption trajectory is a sequence of real-time budget consumption accumulation from the first stage to the current stage. The cumulative trajectory is stored as a time series, where each element corresponds to a decision stage, and the element value is the accumulated budget consumption amount at the end of that stage. The slope of the cumulative trajectory is calculated by dividing the difference in cumulative values ​​between adjacent stages by the stage duration. When the slope exceeds the remaining budget divided by the number of remaining stages, a warning of excessively rapid budget consumption is triggered.

[0097] When mapping the expected long-term conversion revenue to the cumulative budget consumption trajectory onto a dual-objective space, the dual-objective space adopts a two-dimensional planar structure. The horizontal axis represents the cumulative budget consumption amount, and the vertical axis represents the expected long-term conversion revenue. Each deployment action sequence corresponds to a point in the dual-objective space. The horizontal coordinate of the point is the total cumulative budget consumption from the execution of the action sequence to the end of the deployment, and the vertical coordinate is the value function value of the action sequence in the first decision stage. The horizontal coordinate ranges from 0 to the total budget amount, and the vertical coordinate ranges from 0.5 times to 3 times the total budget amount.

[0098] When constructing the dynamic Pareto front, 500 to 2000 candidate action sequences are sampled in the dual-target space. The sampling method combines random sampling and heuristic sampling. Random sampling generates action sequences according to a uniform distribution, while heuristic sampling applies local perturbations based on historical best action sequences. For each candidate action sequence, its coordinates in the dual-target space are calculated, and it is determined whether the point is dominated by other candidate points. Domination is defined as the existence of another point whose x-coordinate is no greater than that of the first point and whose y-coordinate is strictly greater than that of the first point, or whose x-coordinate is strictly less than that of the first point and whose y-coordinate is no less than that of the first point. Points not dominated by any other point constitute the Pareto front set, and the Pareto front forms a piecewise linear curve by connecting the points in the front set.

[0099] A multi-objective evolutionary algorithm framework is employed to iteratively adjust the action sequences deployed across decision stages so that their projection points in the dual-objective space gradually approach the dynamic Pareto front. The algorithm maintains a population of 100 to 300 individuals, each corresponding to a single action sequence. The iterative process includes selection, crossover, and mutation operations. The selection operation samples probabilistically based on the shortest distance an individual has to the Pareto front in the dual-objective space; the closer the distance, the higher the probability of selection. The crossover operation randomly selects a decision stage from two parent individuals, copies the action sequence fragment before that stage from one parent, and the fragment after that stage from the other parent, splicing them together to form a offspring individual. The mutation operation randomly selects several decision stages from the offspring individuals and randomly perturbs the actions in these stages. Perturbation methods include modifying platform selection, adjusting timing allocation weights, and replacing creative content units, with perturbation probabilities set to 0.05 to 0.2. The number of iteration rounds is set to 50 to 200, and the termination condition is reaching the maximum number of iteration rounds or 10 consecutive iterations without change in the Pareto front point set. The algorithm supports parallel computation of individual value functions, with the number of parallel threads set to 4 to 16, and the complete iteration process takes approximately 10 to 60 minutes.

[0100] When selecting a non-dominated solution on a dynamic Pareto front that satisfies the budget constraint boundary conditions and maximizes the expected long-term conversion revenue, the budget constraint boundary conditions are defined as the cumulative budget consumption not exceeding the total budget amount multiplied by the upper limit of budget utilization, with the upper limit of budget utilization set by default to 0.95 to 1.0. Points whose x-coordinates satisfy the budget constraint are selected from the Pareto front set, and the point with the largest y-coordinate among these points is chosen as the non-dominated solution. If multiple points have the same y-coordinate and are all at the maximum value, the point with the smallest x-coordinate is selected to preserve the budget margin.

[0101] When parsing the delivery action sequence corresponding to the non-dominated solution into platform selection strategies, delivery timing allocation schemes, and creative content matching rules for different user groups at each decision-making stage, the platform selection strategy is represented as a three-dimensional data structure, with dimensions of decision-making stage number, user group identifier, and platform identifier set. The delivery timing allocation scheme is also represented as a three-dimensional data structure, with dimensions of decision-making stage number, user group identifier, and time slice weight vector. The creative content matching rules are also represented as a three-dimensional data structure, with dimensions of decision-making stage number, user group identifier, and creative content weight dictionary. All three strategy data structures are stored using nested hash tables, supporting quick lookup of corresponding strategy configurations by decision-making stage number and user group identifier.

[0102] In one optional implementation, mapping the expected long-term conversion revenue to the cumulative trajectory of budget expenditure to a bi-objective space and constructing a dynamic Pareto front includes:

[0103] A dual-objective space is established, with the expected value of long-term conversion revenue as the conversion revenue dimension and the cumulative trajectory of budget consumption as the budget consumption dimension. The expected value of long-term conversion revenue and the cumulative trajectory of budget consumption corresponding to each candidate delivery action sequence in the sequential decision optimization framework are used as coordinate components, and the coordinate components are mapped to the dual-objective space to form a distribution point set.

[0104] For each candidate solution in the distribution point set, count the number of other candidate solutions that are superior to the candidate solution in both the transformation benefit dimension and the budget consumption dimension, quantify their degree of domination, and select candidate solutions with a degree of domination of zero to form a non-dominated solution set. Define the boundary curve formed by the non-dominated solution set in the dual objective space as the dynamic Pareto front.

[0105] Within the bi-objective optimization framework, the multi-dimensional strategy results generated during the sequential decision-making process need to be projected into a quantifiable and comparable space. First, a two-dimensional coordinate system is established, with the horizontal axis representing the budget consumption dimension and the vertical axis representing the conversion revenue dimension. For all candidate deployment action sequences generated by the sequential decision-making optimization framework at the current decision step, each sequence contains a complete deployment plan for the next T time steps. The expected long-term conversion revenue generated after the execution of each sequence is calculated. This expected value is obtained by accumulating the immediate conversion revenue and delayed conversion revenue at each time step, where the delayed conversion revenue is adjusted for time value using a discount factor γ. Simultaneously, the cumulative budget consumption trajectory corresponding to each candidate sequence is tracked, recording the cumulative expenditure from the current moment to the end of the planning period. The expected long-term conversion revenue and the cumulative budget consumption of each candidate sequence are used as coordinate components (c, r), where c represents the cumulative budget consumption and r represents the long-term conversion revenue. This coordinate point is mapped onto the established bi-objective space, forming a distribution set of candidate solutions.

[0106] For any candidate solution i in the distribution set, its relative superiority or inferiority in the dual-objective space needs to be evaluated. All other candidate solutions j in the set are traversed, determining whether candidate solution j is superior to candidate solution i simultaneously in both the conversion benefit and budget consumption dimensions. Specifically, the rule is: when the conversion benefit of candidate solution j is greater than or equal to that of candidate solution i and its budget consumption is less than or equal to that of candidate solution i, and it is strictly superior to candidate solution i in at least one dimension, candidate solution j is considered to dominate candidate solution i. The number of candidate solutions satisfying the dominance relationship is counted and denoted as the degree of dominance of candidate solution i. The degree of dominance reflects the degree of competitive disadvantage of the candidate solution in the current set; a larger value indicates the existence of more superior alternatives.

[0107] All candidate solutions with a degree of dominance of zero are selected. These candidate solutions, for which no other solution in the current point set can simultaneously achieve better performance in both objective dimensions, are classified into the non-dominated solution set. The candidate solutions in the non-dominated solution set represent the optimal trade-offs within the current decision space, reflecting different strategies for balancing budget input and return. The points in the non-dominated solution set are sorted in ascending order of budget consumption, and adjacent points are connected sequentially to form a piecewise linear boundary curve, which is the dynamic Pareto front. Regions on the curve closer to the origin correspond to low-budget, high-efficiency strategies, while regions farther from the origin correspond to high-input, high-return strategies. As the sequential decision-making process moves to the next step, the candidate action sequence is updated, and the mapping and selection process is re-executed, causing the Pareto front to dynamically evolve with the decision-making process, continuously reflecting the optimal strategy boundary set at the current moment.

[0108] In one optional implementation, targeted delivery is executed on the corresponding platform according to the output strategy of the delivery decision model, and multi-dimensional feedback data is collected. The time decay weight calculation rule in the time series analysis process is updated using the multi-dimensional feedback data, including:

[0109] Based on the output strategy of the delivery decision model, targeted delivery is executed on the corresponding platform. During the delivery process, multi-dimensional feedback data is collected, and the time delay distribution characteristics of user response and the time-series fluctuation characteristics of platform bidding are extracted from the multi-dimensional feedback data.

[0110] Based on the time delay distribution characteristics and the time series fluctuation characteristics, an attenuation relationship is constructed, the attenuation relationship is transformed into a weight calculation rule, and the time attenuation weight calculation rule in the time series analysis process is updated using the weight calculation rule.

[0111] After generating the strategy for the ad placement decision model, the system initiates targeted ad placement tasks on the corresponding advertising platforms based on the output platform selection strategy, placement timing allocation plan, and creative content matching rules. During the placement process, feedback data from multiple dimensions, including exposure time, click latency, conversion timestamp, actual bid price, and traffic peak and trough period identifiers, is collected in real time to form a complete ad placement behavior-response data chain.

[0112] Based on the collected multidimensional feedback data, the time delay distribution characteristics of user responses are first extracted. Specifically, this is done by statistically analyzing the time interval from ad exposure to user click and the time span from click to completion of conversion, calculating the distribution density of response rates within different time windows. For example, for multiple time intervals such as 0-5 minutes, 5-30 minutes, 30 minutes-2 hours, and more than 2 hours, the frequency of response events is statistically analyzed to obtain the probability distribution function of time delay. Simultaneously, the differences in response activity among different user groups in different time periods are analyzed to identify the behavioral characteristic dividing points between users who respond quickly and those who respond slowly.

[0113] When extracting the temporal fluctuation characteristics of platform bidding, sampling is performed every hour or every 15 minutes to record data such as average bidding cost, traffic competition intensity indicators, and ad slot supply-demand ratio within each time period. The standard deviation and coefficient of variation of bidding costs are calculated using a sliding window to identify periods of significant price fluctuations. For periodic fluctuations, Fourier transform is used to extract the main periodic components, capturing the differences in bidding patterns between weekdays and weekends, and between daytime and nighttime.

[0114] Based on the extracted time delay distribution and temporal fluctuation characteristics, a decay relationship model is constructed. The probability density function of user response delay is transformed into an interest decay rate parameter. For user groups whose responses are concentrated in short time windows, a steeper exponential decay curve is adopted, with a larger decay coefficient. For groups with longer response time spans, a gentler power-law decay function is used, extending the modeling duration of interest validity. Simultaneously, a correction factor for platform bidding fluctuations is introduced, assigning lower time decay weights to historical behavior data corresponding to high-competition periods, as user behavior during these periods is driven by stronger external stimuli and has a more lasting indicative significance of interest.

[0115] The constructed decay relationship is transformed into an executable weight calculation rule. The basic form of the time decay weight is set as a function of the time difference between the historical event and the current time. This is combined with user group response latency characteristic parameters and platform bidding fluctuation correction coefficients to form a piecewise function or a parameterized continuous function. For example, for a fast-responding young user group, the weight calculation rule adopts... The form is used, where the decay rate λ is determined based on the reciprocal of the average response delay of the population; for the middle-aged and elderly population with slower response, a more specific approach is adopted. The power-law form of the response delay distribution is obtained by fitting the tail features of the response delay distribution.

[0116] The time decay weight of each historical behavior record in the time series analysis is recalculated using the updated weight calculation rules. This new weight is then applied to the generation of the user interest representation vector, giving higher weight to recent behaviors and historical behaviors that conform to the current platform environment, while reasonably mitigating the impact of outdated or abnormal bidding environment-generated behavior data. The weight update trigger frequency is set to once every 24 hours or when the accumulated new feedback samples reach a set threshold, ensuring a balance between the model's tracking accuracy of user interest evolution and computational resource consumption.

[0117] A second aspect of the present invention provides a cross-platform user behavior analysis and media information matching system, comprising:

[0118] The data fusion unit is used to acquire user behavior data from multiple heterogeneous platforms, perform time-series analysis on the user behavior data, capture the evolution of user behavior patterns under different time windows, identify the phased transfer characteristics and decay patterns of user interests, generate dynamic interest representation vectors with time decay weights, and map the heterogeneous behavior features of different platforms to a unified semantic space to obtain the distribution of potential user interests that integrates information from multiple platforms.

[0119] The profile building unit is used to build multi-granularity user group profiles based on the matching relationship between the distribution of potential user interests and the semantic features of platform content;

[0120] The decision optimization unit is used to establish a campaign decision model. Based on the campaign decision model, the multi-granular user group profile, platform bidding environment, and time period traffic distribution are used as state inputs. Through the sequential decision optimization framework, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints, the unit dynamically generates platform selection strategies, campaign timing allocation schemes, and creative content matching rules for different user groups. The decision strategy is continuously evolved based on the cumulative feedback of historical campaign performance.

[0121] The closed-loop execution unit is used to execute targeted delivery on the corresponding platform according to the output strategy of the delivery decision model, collect multi-dimensional feedback data, update the time decay weight calculation rules in the time series analysis process using the multi-dimensional feedback data, and adjust the strategy generation parameters of the delivery decision model to form a two-way closed loop of user understanding and delivery optimization.

[0122] A third aspect of the present invention provides an electronic device, comprising:

[0123] processor;

[0124] Memory used to store processor-executable instructions;

[0125] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0126] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0127] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-platform user behavior analysis and media information matching method, characterized in that, include: User behavior data from multiple heterogeneous platforms is acquired, and time-series analysis is performed on the user behavior data. By capturing the evolution of user behavior patterns under different time windows, the phased transfer characteristics and decay patterns of user interests are identified, a dynamic interest representation vector with time decay weight is generated, and the heterogeneous behavior features of different platforms are mapped to a unified semantic space to obtain the distribution of user potential interests that integrates information from multiple platforms. Based on the matching relationship between the user's potential interest distribution and the semantic features of the platform content, a multi-granularity user group profile is constructed. Establish a campaign decision model. Based on the campaign decision model, take the multi-granular user group profile, platform bidding environment, and time period traffic distribution as state inputs. Through a sequential decision optimization framework, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints, dynamically generate platform selection strategies, campaign timing allocation schemes, and creative content matching rules for different user groups. And continuously evolve the decision strategy based on the cumulative feedback of historical campaign performance. Based on the output strategy of the aforementioned delivery decision model, targeted delivery is executed on the corresponding platform, and multi-dimensional feedback data is collected. The time decay weight calculation rules in the time series analysis process are updated using the multi-dimensional feedback data, and the strategy generation parameters of the delivery decision model are adjusted at the same time, forming a two-way closed loop of user understanding and delivery optimization.

2. The method according to claim 1, characterized in that, Time-series analysis is performed on the user behavior data. By capturing the evolution of user behavior patterns in different time windows, the phased shift characteristics and decay patterns of user interests are identified, and a dynamic interest representation vector with time decay weights is generated, including: User behavior data is divided into multiple consecutive time windows according to the time dimension, and behavioral feature vectors are extracted for the user interaction sequence within each time window. Establish a behavioral evolution correlation map between time windows, identify the migration trajectory of user interests between different topic domains by calculating the semantic distance and distribution offset of behavioral feature vectors between adjacent time windows, and determine the stage boundary of interest transfer based on the persistence and retrospectiveness of the migration trajectory, so as to obtain the stage transfer feature identifier that divides the interest evolution stage. For each interest evolution stage in the phased transition feature identifier, a two-factor decay function is constructed based on the time span between the interest evolution stage and the current time and the fluctuation amplitude of the behavioral feature vector within the interest evolution stage. The time span is mapped to a time distance decay factor and the behavioral fluctuation amplitude is mapped to a stability decay factor according to the two-factor decay function. The final decay coefficient is calculated by fusing the time distance decay factor and the stability decay factor to generate a time decay weight sequence that matches the characteristics of the interest evolution stage. The behavioral feature vector is fused with the time decay weight sequence at corresponding positions to generate a dynamic interest representation vector with time decay weights.

3. The method according to claim 1, characterized in that, Based on the matching relationship between the user's potential interest distribution and the semantic features of the platform content, a multi-granularity user group profile is constructed, including: The distribution of users’ potential interests is mapped to the semantic feature space of platform content. By constructing an interest-content matching metric function, the response strength between each interest dimension in the user’s interest distribution and the semantic features of platform content is calculated. The response strength reflects both the sensitivity of the interest dimension to a specific content type and the triggering ability of the content feature to activate the interest, thus obtaining an interest-content matching relationship matrix that characterizes the individual content consumption preferences of users. Based on the interest-content matching relationship matrix, cross-user response pattern analysis is performed. By identifying the distribution pattern of response intensity and activation path similarity of different users in the interest-content matching relationship matrix, users with similar response patterns are aggregated into user groups. For each user group, common response patterns and specific response patterns are extracted from the matching relationship matrix of its members to generate group-level feature identifiers that distinguish different groups. By analyzing the overlapping areas and evolution trends of response patterns among the group-level feature identifiers, and by tracking the change trajectory of user response patterns in different groups, a cross-group association map describing the migration patterns of users between groups is constructed. The interest-content matching matrix, the group-level feature identifier, and the cross-group association graph are combined to form a multi-granularity user group profile.

4. The method according to claim 3, characterized in that, Analyzing the overlapping areas and evolution trends of response patterns among the group-level feature identifiers, and by tracking the changing trajectories of user response patterns across different groups, a cross-group association map describing the migration patterns of users between groups is constructed, including: Cross-group response pattern spatial projection is performed on group-level feature identifiers. By constructing a unified response pattern representation space and mapping the response patterns of different groups to the unified response pattern representation space, the spatial overlap region and separation boundary of response patterns between groups are calculated. Group pairs with continuous response features and group pairs with broken response are identified, and the topological structure of the overlapping region that quantifies the similarity and difference of responses between groups is obtained. Based on the overlapping region topology, the user's group affiliation evolution trajectory in the time series is tracked. By analyzing the temporal movement path of the user response pattern in the unified response pattern representation space, the migration trajectory of the user from the source group response region to the target group response region is identified. Based on the spatial span, migration speed and trajectory density distribution of the migration trajectory, the directional features of user migration between groups, migration concentration and response pattern change threshold are extracted. Using the population pairs in the overlapping region topology as graph nodes, the directional features of the migration trajectory and the migration concentration as directed connection edge attributes between nodes, and the response pattern change threshold as the edge activation condition, a cross-population association graph is constructed.

5. The method according to claim 1, characterized in that, By employing a sequential decision optimization framework aimed at maximizing long-term conversion benefits and achieving Pareto optimality within budget constraints, the framework dynamically generates platform selection strategies, ad placement timing allocation schemes, and creative content matching rules for different user groups, including: A sequential decision optimization framework is established, and the campaign process is divided into multiple consecutive decision stages according to the sequential decision optimization framework. In each decision stage, a state-action-value mapping relationship is constructed to predict the fluctuation of the platform bidding environment and the migration of traffic distribution triggered by the campaign action targeting a specific user group in the current stage. By retrospectively tracing the cumulative impact of the conversion revenue contribution and budget consumption burden of future stages on the value of the action in the current stage, the long-term conversion revenue expectation and budget consumption cumulative trajectory of the campaign action in the current stage are quantified. The expected long-term conversion revenue is mapped to a dual-objective space along with the cumulative budget consumption trajectory, and a dynamic Pareto front is constructed. By iteratively adjusting the sequence of delivery actions across decision-making stages, the projection point of the action in the dual-objective space gradually approaches the dynamic Pareto front. On the dynamic Pareto front, a non-dominated solution that satisfies the budget constraint boundary conditions and maximizes the expected long-term conversion revenue is selected. The delivery action sequence corresponding to the non-dominated solution is then analyzed into platform selection strategies, delivery timing allocation schemes, and creative content matching rules for different user groups at each decision-making stage.

6. The method according to claim 5, characterized in that, Mapping the expected long-term conversion revenue and the cumulative trajectory of budget expenditure to a dual-objective space and constructing a dynamic Pareto front includes: A dual-objective space is established, with the expected value of long-term conversion revenue as the conversion revenue dimension and the cumulative trajectory of budget consumption as the budget consumption dimension. The expected value of long-term conversion revenue and the cumulative trajectory of budget consumption corresponding to each candidate delivery action sequence in the sequential decision optimization framework are used as coordinate components, and the coordinate components are mapped to the dual-objective space to form a distribution point set. For each candidate solution in the distribution point set, count the number of other candidate solutions that are superior to the candidate solution in both the transformation benefit dimension and the budget consumption dimension, quantify their degree of domination, and select candidate solutions with a degree of domination of zero to form a non-dominated solution set. Define the boundary curve formed by the non-dominated solution set in the dual objective space as the dynamic Pareto front.

7. The method according to claim 1, characterized in that, Based on the output strategy of the aforementioned delivery decision model, targeted delivery is executed on the corresponding platform, and multi-dimensional feedback data is collected. The time decay weight calculation rules in the time series analysis process are updated using the multi-dimensional feedback data, including: Based on the output strategy of the delivery decision model, targeted delivery is executed on the corresponding platform. During the delivery process, multi-dimensional feedback data is collected, and the time delay distribution characteristics of user response and the time-series fluctuation characteristics of platform bidding are extracted from the multi-dimensional feedback data. Based on the time delay distribution characteristics and the time series fluctuation characteristics, an attenuation relationship is constructed, the attenuation relationship is transformed into a weight calculation rule, and the time attenuation weight calculation rule in the time series analysis process is updated using the weight calculation rule.

8. A cross-platform user behavior analysis and media information matching system, used to implement the method of any one of claims 1-7, characterized in that, include: The data fusion unit is used to acquire user behavior data from multiple heterogeneous platforms, perform time-series analysis on the user behavior data, capture the evolution of user behavior patterns under different time windows, identify the phased transfer characteristics and decay patterns of user interests, generate dynamic interest representation vectors with time decay weights, and map the heterogeneous behavior features of different platforms to a unified semantic space to obtain the distribution of potential user interests that integrates information from multiple platforms. The profile building unit is used to build multi-granularity user group profiles based on the matching relationship between the distribution of potential user interests and the semantic features of platform content; The decision optimization unit is used to establish a campaign decision model. Based on the campaign decision model, the multi-granular user group profile, platform bidding environment, and time period traffic distribution are used as state inputs. Through the sequential decision optimization framework, with the goal of maximizing long-term conversion benefits and Pareto optimality under budget constraints, the unit dynamically generates platform selection strategies, campaign timing allocation schemes, and creative content matching rules for different user groups. The decision strategy is continuously evolved based on the cumulative feedback of historical campaign performance. The closed-loop execution unit is used to execute targeted delivery on the corresponding platform according to the output strategy of the delivery decision model, collect multi-dimensional feedback data, update the time decay weight calculation rules in the time series analysis process using the multi-dimensional feedback data, and adjust the strategy generation parameters of the delivery decision model to form a two-way closed loop of user understanding and delivery optimization.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images