AI deep analysis method and device for multi-platform advertisement putting effect data

Abstract

The invention provides an AI deep analysis method and device for multi-platform advertisement putting effect data, and the method comprises the steps: carrying out the incremental updating of the advertisement putting effect data of each putting platform based on a preset time-space attenuation factor, and generating a conversion behavior graph; performing association analysis on the dynamic conversion behavior map to obtain a conversion behavior pattern; the advertisement effect data is divided into associated data blocks according to the conversion behavior modes, each associated data block corresponds to a unique conversion behavior mode, and the conversion behavior modes corresponding to different associated data blocks meet orthogonality constraints; and inputting the associated data blocks with the same conversion behavior pattern into a preset analysis model to obtain a putting contribution degree score of each putting platform.

Description

Technical Field

[0001] This application relates to the field of advertising technology, and in particular to an AI-based deep analysis method and apparatus for multi-platform advertising performance data. Background Technology

[0002] Currently, businesses typically advertise on multiple platforms simultaneously (such as search engines, social media, and e-commerce platforms), making user conversion paths extremely complex. Traditional attribution analysis methods have significant limitations: first-click and last-click attribution are overly simplistic and fail to reflect the synergistic effects of multiple touchpoints; linear attribution models ignore the actual weight differences between different platforms in the conversion chain; and rule-based attribution methods lack dynamic adaptability. Existing technologies struggle to handle the time decay of cross-platform data, dynamic changes in user behavior, and the interactive effects between platforms. Summary of the Invention

[0003] This application provides an AI-based deep analysis method and apparatus for multi-platform advertising performance data, which is used to comprehensively analyze the advertising performance of different platforms to assist in subsequent advertising investment strategies.

[0004] In a first aspect, embodiments of this application provide an AI-based deep analysis method for multi-platform advertising performance data, the method comprising: Based on a preset spatiotemporal decay factor, the advertising performance data of each advertising platform is incrementally updated to generate a conversion behavior map. Correlation analysis was performed on the dynamic transformation behavior map to obtain the transformation behavior pattern; The advertising performance data is divided into associated data blocks according to the conversion behavior pattern. Each associated data block corresponds to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different associated data blocks satisfy the orthogonality constraint. By inputting associated data blocks with the same conversion behavior pattern into a preset analysis model, a delivery contribution score for each delivery platform is obtained.

[0005] In some embodiments, before incrementally updating the advertising performance data for each platform based on a preset spatiotemporal decay factor, the method further includes: A preset spatiotemporal decay factor is generated based on user behavior timestamps and platform environment change rate. The platform environment change rate includes: platform traffic fluctuation coefficient and competitive advertising density index.

[0006] In some embodiments, the step of performing correlation analysis on the dynamic conversion behavior map to obtain conversion behavior patterns includes: Extract the node feature vectors and edge weight information from the transformation behavior graph; Based on the node feature vectors and the edge weight information, identify key transformation paths; The key conversion paths are summarized into patterns to generate conversion behavior patterns.

[0007] In some embodiments, dividing the advertising performance data into associated data blocks based on the conversion behavior pattern includes: Based on the conversion behavior pattern, cluster analysis is performed on the advertising effect data to generate user behavior clusters. The similarity of the feature vectors of different user behavior clusters satisfies the preset orthogonality constraint. Assign a unique behavior pattern tag to each user behavior cluster and establish a binding relationship table between the tag and the advertising performance data; Based on the binding relationship table, the advertising performance data is divided into multiple related data blocks.

[0008] In some embodiments, inputting associated data blocks with the same conversion behavior pattern into a preset analysis model to obtain a campaign contribution score includes: The associated data block is input into a pre-set parallelized graph neural network to obtain the projection feature representation; Based on the aforementioned delivery feature representation, an adversarial evaluation model is constructed to calculate the delivery contribution score for each delivery platform.

[0009] In some embodiments, the step of clustering the advertising performance data based on the conversion behavior pattern to generate user behavior clusters, wherein the feature vector similarity of different user behavior clusters satisfies a preset orthogonality constraint, includes: Extract user behavior features from the advertising performance data; Based on the path topology of the conversion behavior pattern, the user behavior features are segmented to generate an initial set of behavior clusters; Calculate the centroid cosine similarity of different behavior clusters in the initial behavior cluster set to generate a similarity matrix; The similarity matrix is ​​adjusted by an orthogonal projection algorithm so that the centroid cosine similarity is less than or equal to a preset threshold, thereby obtaining a set of user behavior clusters.

[0010] Secondly, embodiments of this application provide an AI deep analysis device for multi-platform advertising performance data, the AI ​​deep analysis device for multi-platform advertising performance data includes: The graph generation module is used to incrementally update the advertising performance data of each advertising platform based on a preset spatiotemporal decay factor, and generate a conversion behavior graph. The behavior analysis module is used to perform correlation analysis on the dynamic conversion behavior map to obtain conversion behavior patterns; A data segmentation module is used to divide the advertising performance data into associated data blocks according to the conversion behavior pattern. Each associated data block corresponds to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different associated data blocks satisfy orthogonality constraints. The scoring and analysis module is used to input associated data blocks with the same conversion behavior pattern into a preset analysis model to obtain the delivery contribution score of each delivery platform.

[0011] This application provides an AI-powered deep analysis method for multi-platform advertising performance data. The method includes: incrementally updating the advertising performance data for each platform based on a preset spatiotemporal decay factor to generate a conversion behavior graph; performing correlation analysis on the dynamic conversion behavior graph to obtain conversion behavior patterns; dividing the advertising performance data into related data blocks according to the conversion behavior patterns, with each related data block corresponding to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different related data blocks satisfying orthogonality constraints; and inputting related data blocks with the same conversion behavior pattern into a preset analysis model to obtain a contribution score for each platform. In this method, by introducing a spatiotemporal decay factor for incremental updates and constructing a conversion behavior graph, combined with orthogonal constraint-constrained related data block division and a dedicated analysis model, it is possible to dynamically capture changes in user behavior, accurately identify complex conversion path patterns, and avoid data overlap bias. This enables precise quantitative evaluation of the contribution of advertising across multiple platforms, providing a scientific and reliable decision-making basis for optimizing advertising strategies. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic flowchart illustrating an AI-based deep analysis method for multi-platform advertising performance data provided in this application embodiment; Figure 2 This is a schematic block diagram of an AI deep analysis device for multi-platform advertising performance data provided in an embodiment of this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described below with reference to the accompanying drawings.

[0015] The terms "first" and "second," etc., used in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0016] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0017] It should be understood that in this application, "at least one (item)" means one or more, "more than one" means two or more, "at least two (items)" means two or three or more, and "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the related objects before and after are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0018] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating an AI-based deep analysis method for multi-platform advertising performance data provided in an embodiment of this application. Figure 1 As shown, the specific steps of this AI-based deep analysis method for multi-platform advertising performance data include: S101-S104.

[0019] S101. Based on the preset spatiotemporal decay factor, incrementally update the advertising performance data of each advertising platform to generate a conversion behavior map.

[0020] For example, the original advertising performance data from each advertising platform is obtained, including click timestamps, conversion events, user identifiers, and platform identifiers. The time difference between the user behavior timestamp and the current moment is calculated, and the platform traffic fluctuation coefficient and competitive ad density index are extracted as components of the platform environment change rate. The time difference and the platform environment change rate are weighted and combined to generate a spatiotemporal decay factor value for each data record. The spatiotemporal decay factor is applied to key indicators such as conversion weight and click value in the advertising performance data, and the values ​​are updated. A mapping relationship between user nodes and ad touchpoint nodes is established, and the updated data is converted into a graph structure representation. An attribute vector is assigned to each node in the graph, containing platform type, touch time, and decayed performance indicators. Directed edges are established between users and ad touchpoints, with edge weights reflecting the touch intensity and time series relationship. The same decay factor calculation and node creation process is performed on newly added advertising data, and incremental updates to the conversion behavior graph are achieved through dynamic edge addition and weight adjustment. The time series integrity of the graph is maintained to ensure the correct representation of the user conversion path in the graph structure.

[0021] S102. Perform correlation analysis on the dynamic conversion behavior map to obtain the conversion behavior pattern.

[0022] For example, the degree centrality, proximity centrality, and betweenness centrality of each node are extracted from the conversion behavior graph as the basic components of the node feature vector. The platform affiliation, temporal location, and conversion contribution value of each node are calculated, and these attributes are encoded into vector form and added to the feature vector. All edge connections in the graph are traversed, edge weight information is extracted, and the directional coefficient of the edges is calculated. A path search algorithm is used to identify the complete user trajectory from initial contact to conversion, and paths with a total weight exceeding a preset threshold are selected as key conversion paths. Path length statistics, platform jump frequency calculations, and conversion time span analysis are performed on key conversion paths. Conversion trajectories with similar path characteristics are grouped into the same category, and reusable conversion behavior pattern templates are generated through path pattern abstraction. A unique identifier is assigned to each conversion behavior pattern, and the platform sequence involved in the pattern, typical conversion duration, and average conversion value are recorded. All user paths in the dynamic conversion behavior graph are matched with the identified conversion behavior patterns to complete the association analysis process.

[0023] S103. Divide the advertising performance data into related data blocks according to the conversion behavior pattern. Each related data block corresponds to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different related data blocks satisfy the orthogonality constraint.

[0024] For example, user behavior features are extracted from advertising performance data, including platform access frequency, dwell time, click response speed, and conversion completion rate. Based on the path topology information in the conversion behavior pattern, user behavior features are grouped and segmented according to the path node positions. Clustering operations are performed on the segmented feature data to generate an initial set of behavior clusters reflecting different user behavior preferences. The centroid coordinates of each behavior cluster are calculated, and the similarity value between the centroids of different behavior clusters is obtained using the cosine similarity calculation formula. The centroid cosine similarity values ​​are filled into the corresponding positions in the similarity matrix, and it is checked whether there are any behavior cluster pairs in the matrix with similarity exceeding a preset threshold. An orthogonal projection algorithm is applied to adjust the coordinates of high-similarity behavior clusters, and the centroid cosine similarity is reduced to within the threshold range through iterative calculation. A unique behavior pattern label is assigned to each adjusted user behavior cluster, and a corresponding binding relationship table between label numbers and advertising performance data records is established. Based on the label grouping information in the binding relationship table, the original advertising performance data is divided according to label categories, generating multiple independent associated data blocks. The conversion behavior patterns corresponding to different associated data blocks are verified to meet the orthogonality constraint requirements.

[0025] S104. Input the associated data blocks with the same conversion behavior pattern into the preset analysis model to obtain the delivery contribution score of each delivery platform.

[0026] For example, each associated data block is converted into an input format suitable for parallel graph neural network processing, including the construction of node feature matrices and adjacency matrices. Key parameters such as the number of layers, number of hidden units, and activation function type of the parallel graph neural network are configured. Feature extraction and dimensionality transformation of node information in the associated data blocks are performed through graph convolution operations to obtain a delivery feature representation vector containing platform interaction information. The feature extraction process is executed in parallel for multiple associated data blocks with the same conversion behavior pattern to improve computational efficiency. Generator and discriminator components of an adversarial evaluation model are constructed based on the delivery feature representation. The generator receives the delivery feature representation as input and outputs a predicted contribution distribution for each delivery platform. The discriminator is responsible for distinguishing between the true contribution distribution and the predicted distribution output by the generator. The parameter settings of the generator and discriminator are optimized through an adversarial training process, enabling the generator to accurately predict the contribution of each delivery platform under a specific conversion behavior pattern. The delivery contribution score values ​​for each delivery platform are extracted from the trained generator to complete the quantitative evaluation of the multi-platform advertising performance.

[0027] To more clearly illustrate the technical solution of this application, the technical solution of this application will be described below through specific embodiments. It should be noted that the specific embodiments are used to expand the description of the technical solution of this application, and are not intended to limit this application.

[0028] In some embodiments, before incrementally updating the advertising performance data of each advertising platform based on a preset spatiotemporal decay factor, the method further includes: generating a preset spatiotemporal decay factor based on user behavior timestamps and platform environment change rate, wherein the platform environment change rate includes: platform traffic fluctuation coefficient and competitive advertising density index.

[0029] For example, in the process of generating a preset spatiotemporal decay factor based on user behavior timestamps and platform environment change rate, the user behavior timestamp is constructed by analyzing the specific time points of behaviors such as ad clicks, browsing, and conversions to create a time series feature vector. This vector records the user's activity change patterns and behavior frequency distribution patterns in different time periods. The calculation of the platform environment change rate involves real-time monitoring of the platform traffic fluctuation coefficient. By collecting hourly access data from each advertising platform and calculating the ratio of its standard deviation to the mean, traffic stability is quantified. When the ratio exceeds a set range, a dynamic adjustment mechanism for the decay factor is triggered. The competitive ad density index assesses the intensity of competition by statistically analyzing the number of similar ads placed and the frequency of display on each platform, and uses a density function to calculate the distribution density value of competitive ads per unit time. The preset spatiotemporal decay factor uses a multiple regression algorithm to fuse timestamp features and loop data. A weighted average method is used to combine the decay weights of the time dimension with the environmental impact factors of the spatial dimension to generate personalized decay parameters applicable to different advertising platforms. These parameters can accurately reflect the natural decay trend of advertising effectiveness in the spatiotemporal dimension.

[0030] In the incremental update process of advertising performance data, the generation of the spatiotemporal decay factor needs to comprehensively consider two core dimensions: user behavior timestamps and platform environment change rate. User behavior timestamps reflect the changing patterns of user activity at different points in time. By analyzing the time distribution characteristics of user clicks, browsing, and purchases, the periodic patterns and decay trends of user behavior can be identified. For example, the click-through rate of users on weekday evenings from 8-10 pm is 35% higher than at noon, while the conversion rate on weekends is 15% lower than on weekdays. The platform environment change rate is another important factor. The platform traffic fluctuation coefficient quantifies platform stability by monitoring the real-time traffic changes of each advertising platform. When the traffic fluctuation of a platform exceeds 20% during a specific period, the corresponding decay weight will be adjusted accordingly. The competitive ad density index assesses the intensity of competition by analyzing the density of similar ads on various platforms. When the competitive ad density exceeds 1.5 times the average, the expected advertising performance on that platform will be affected by a corresponding decay. Based on this multi-dimensional data, the preset spatiotemporal decay factor is generated using a weighted fusion algorithm, with the time dimension accounting for 40% and the spatial dimension accounting for 60%. By dynamically adjusting these weight parameters, the decay pattern of advertising effectiveness on different platforms under specific spatiotemporal conditions can be more accurately reflected.

[0031] In some embodiments, correlation analysis is performed on the dynamic conversion behavior graph to obtain conversion behavior patterns, including: extracting node feature vectors and edge weight information from the conversion behavior graph; identifying key conversion paths based on node feature vectors and edge weight information; and summarizing the key conversion paths to generate conversion behavior patterns.

[0032] For example, the process of performing correlation analysis on a dynamic conversion behavior graph includes extracting node feature vectors and edge weight information from the graph structure. Node feature vectors carry multi-dimensional data such as user attribute information, historical behavior records, and preference tags. These heterogeneous data are converted into a unified numerical vector representation using a feature encoding algorithm. Edge weight information reflects the correlation strength between different user behavior nodes. Mutual information calculation methods are used to quantify the dependencies between nodes and assign corresponding weight values. The identification of key conversion paths is based on graph traversal algorithms and path scoring mechanisms. Key conversion sequences are determined by calculating the path combination with the highest cumulative weight value among all possible paths from the starting node to the target node. The pattern induction process uses cluster analysis technology to classify and organize the identified key conversion paths. Based on similarity indicators across multiple dimensions such as path length, node type distribution, and weight distribution characteristics, paths are categorized into different behavior pattern categories. Each category represents a typical user conversion behavior pattern, forming a set of conversion behavior patterns covering multiple user decision-making modes.

[0033] The correlation analysis process of the dynamic conversion behavior graph involves a deep analysis of the complex network structure. The node feature vectors contain multi-dimensional attributes such as user demographics, historical behavioral preferences, and device characteristics. Each node's feature vector has 128 dimensions, covering key information such as age, gender, geographic location, spending power, and interest tags. Edge weights quantify the correlation strength between different user behaviors. Edge weights are determined by calculating the conversion probability at each stage, from browsing to clicking, from clicking to adding to cart, and from adding to cart to purchasing. The weights range from 0 to 1, with edges having a weight greater than 0.7 identified as strong correlations. The identification of key conversion paths uses a graph traversal algorithm combined with path weight accumulation. When the total weight of a path exceeds a preset threshold of 2.5 and contains at least three key nodes, the path is marked as a key conversion path. The pattern induction process uses cluster analysis to classify and organize the identified key conversion paths. By analyzing the similarity of the path's topological structure, the distribution characteristics of node types, and the weight distribution pattern, conversion paths with similar characteristics are grouped into the same conversion behavior pattern, generating a complete classification system that includes various typical conversion behavior patterns such as impulsive purchase, rational comparison, and social influence.

[0034] In some embodiments, dividing advertising performance data into associated data blocks based on conversion behavior patterns includes: performing cluster analysis on advertising performance data based on conversion behavior patterns to generate user behavior clusters, wherein the similarity of feature vectors of different user behavior clusters satisfies a preset orthogonality constraint; assigning a unique behavior pattern label to each user behavior cluster and establishing a binding relationship table between the label and advertising performance data; and dividing the advertising performance data into multiple associated data blocks based on the binding relationship table.

[0035] For example, the advertising performance data segmentation process guided by conversion behavior patterns adopts a multi-level clustering analysis strategy. When performing deep clustering analysis on advertising performance data based on conversion behavior patterns, it is necessary to consider information from multiple dimensions such as user behavior sequence characteristics, time interval distribution, and platform preferences. During the generation of user behavior clusters, each cluster contains a group of users with similar conversion paths and behavioral characteristics. The behavioral similarity of users within a cluster is calculated using cosine similarity, requiring that the average similarity of users within the same cluster be no less than 0.8. The feature vector similarity between different user behavior clusters needs to meet a preset orthogonality constraint, that is, the cosine similarity between the centroid vectors of any two different behavior clusters should be less than 0.3, ensuring that each behavior cluster has significant distinguishability and independence. The assignment of behavior pattern labels adopts semantic naming rules, such as "high-frequency browsing - quick decision-making type" and "multi-platform price comparison - delayed purchase type," with each label corresponding to a unique numerical code for easy computer processing. The binding relationship table establishes a mapping relationship between tags and advertising performance data. The table records key information such as behavioral pattern tags, user ID, delivery platform, and timestamps corresponding to each piece of advertising performance data. Based on this binding relationship table, advertising performance data is systematically divided into multiple related data blocks, and each data block contains all relevant data records for the same behavioral pattern tag.

[0036] In some embodiments, related data blocks with the same conversion behavior pattern are input into a preset analysis model to obtain a delivery contribution score, including: inputting the related data blocks into a preset parallelized graph neural network to obtain a delivery feature representation; and constructing an adversarial evaluation model based on the delivery feature representation to calculate the delivery contribution score for each delivery platform.

[0037] When related data blocks with the same conversion behavior patterns are processed by a pre-defined analysis model, a parallelized graph neural network architecture is used to extract deep-level delivery feature representations. The parallelized graph neural network contains multiple parallel graph convolutional layers, each responsible for processing specific types of feature information, such as user attribute features, behavioral sequence features, and platform features. The hidden layer dimension is set to 256 dimensions, and through a combination of 3 graph convolutional operations and 2 fully connected layers, a 128-dimensional delivery feature representation vector is output. This vector comprehensively encodes multi-dimensional information such as user behavior patterns, platform characteristics, and time factors, fully reflecting the impact of different delivery platforms on user conversion behavior. The adversarial evaluation model is built based on the idea of ​​generative adversarial networks (GANs), containing two sub-networks: a generator and a discriminator. The generator is responsible for generating predicted delivery performance scores based on the delivery feature representations, while the discriminator is responsible for distinguishing between the actual delivery performance and the generator's predicted performance. Through adversarial training of the two networks, the model learns a more accurate and robust scoring mechanism. The contribution score for each advertising platform is calculated by comprehensively considering factors such as the platform's performance differences under different user behavior patterns, the synergistic effect with other platforms, and its importance weight in the entire conversion path. The score ranges from 0 to 100, with a higher score indicating a greater contribution from the platform.

[0038] In some embodiments, cluster analysis is performed on advertising performance data based on conversion behavior patterns to generate user behavior clusters. The similarity of feature vectors of different user behavior clusters satisfies a preset orthogonality constraint. This includes: extracting user behavior features from advertising performance data; segmenting user behavior features based on the path topology of conversion behavior patterns to generate an initial set of behavior clusters; calculating the centroid cosine similarity of different behavior clusters in the initial set of behavior clusters to generate a similarity matrix; and adjusting the similarity matrix using an orthogonal projection algorithm so that the centroid cosine similarity is less than or equal to a preset threshold to obtain a set of user behavior clusters.

[0039] For example, the process of clustering advertising performance data based on conversion behavior patterns to generate user behavior clusters involves the deep extraction and analysis of user behavior features. These features include multi-dimensional behavioral indicators such as click frequency, browsing duration, conversion path, and purchase preferences. These features are converted into computable numerical vectors using feature engineering methods. The path topology of conversion behavior patterns provides theoretical guidance for the segmentation of user behavior features. User behavior features are structured and grouped based on the node connections and topological features of different conversion paths. The initial set of behavior clusters is generated using a basic clustering algorithm, containing preliminary results of user group segmentation based on behavioral similarity, providing a basic data structure for subsequent optimization and adjustment. The calculation of centroid cosine similarity uses the cosine distance metric in vector space. The cosine of the angle between the centroid vectors of different behavior clusters is used to quantify the similarity between clusters. The generated similarity matrix records the similarity values ​​between all cluster pairs. The orthogonal projection algorithm adjusts the similarity matrix through mathematical transformation and iterative optimization, so that the centroid cosine similarity can meet the constraint condition of being less than or equal to a preset threshold. After processing by the algorithm, a set of user behavior clusters that meet the orthogonality requirement is obtained.

[0040] The extraction of user behavior features from advertising performance data encompasses the complete behavioral trajectory of users across various advertising platforms, including multi-dimensional behavioral indicators such as click count, dwell time, page navigation path, purchase amount, and review behavior. Each user's behavioral feature vector contains more than 50 different feature dimensions. The path topology of conversion behavior patterns provides a theoretical basis for the segmentation of user behavior features. By analyzing the node connection relationships and weight distribution characteristics of different conversion paths, user behavior features are grouped and segmented according to the path structure. The generated initial behavior cluster set contains user group segmentation results based on path similarity. The centroid cosine similarity calculation of different behavior clusters in the initial behavior cluster set adopts the cosine distance metric method in high-dimensional vector space. The similarity between clusters is quantified by calculating the cosine value of the angle between the centroid vectors of each behavior cluster. The generated similarity matrix is ​​a symmetric matrix, and each element in the matrix represents the degree of similarity between the corresponding two behavior clusters. The application of the orthogonal projection algorithm aims to adjust the numerical distribution in the similarity matrix through mathematical transformation, so that the centroid cosine similarity between different behavior clusters can meet the constraint condition of being less than or equal to a preset threshold of 0.25. The algorithm gradually adjusts the centroid position and feature weight allocation of each behavior cluster through an iterative optimization process. After multiple rounds of iterative calculation, a set of user behavior clusters that meet the orthogonality constraint condition is obtained, ensuring that different behavior clusters have good distinguishability and independence.

[0041] In some embodiments, segmenting the user behavior features based on the path topology of the conversion behavior pattern to generate an initial behavior cluster set includes: constructing an adjacency matrix of the conversion behavior pattern, where the element values ​​of the adjacency matrix represent the connection weights between path nodes; calculating the connectivity components of the path topology based on the adjacency matrix, and grouping path nodes with the same connectivity components into the same topological substructure; performing dimensional mapping on the user behavior features according to the node distribution characteristics of the topological substructure to generate feature subsets corresponding to the topological substructure; calculating the variance contribution rate of the user behavior features within each feature subset, and using feature subsets with variance contribution rates greater than a preset threshold as effective segmentation units; and clustering the user behavior features based on the effective segmentation units to generate an initial behavior cluster set.

[0042] For example, in the process of segmenting user behavior features based on the path topology structure of conversion behavior patterns, the adjacency matrix is ​​constructed by analyzing the direct connection relationships and indirect association strengths between various path nodes in the conversion behavior pattern. The adjacency matrix adopts a sparse matrix storage format to improve computational efficiency. The element values ​​in the matrix are determined by calculating the product of the conversion probability and the time decay factor between path nodes to determine the numerical value of the connection weight. The calculation of connectivity components uses a depth-first search algorithm to traverse the non-zero elements of the adjacency matrix. Different path topology regions are divided by identifying the sets of interconnected nodes in the matrix. Each connectivity component represents an independent user behavior conversion subgraph. The formation of the topology substructure is based on the partitioning results of the connectivity components. Path nodes with the same connectivity component identifier are organized into a hierarchical tree structure. The node distribution characteristics of the topology substructure include topological parameters such as the depth, width, and node density of the substructure. The dimensionality mapping process projects the high-dimensional user behavior feature vector onto the low-dimensional feature space corresponding to the substructure according to the spatial distribution characteristics of the topology substructure. The generated feature subset maintains the main change patterns of the original behavior features within the corresponding topology region. The variance contribution rate is calculated by evaluating the explanatory power of each feature subset on the overall behavioral feature variability using principal component analysis. When the variance contribution rate of a feature subset exceeds a set threshold, the subset is marked as a valid segmentation unit containing important behavioral pattern information. The clustering operation performs an initial division of user groups based on the similarity measure of feature vectors within the valid segmentation units. Cluster centers and cluster boundaries are determined by calculating the Euclidean distance and cosine similarity between feature vectors. The generated initial behavioral cluster set provides the basic data structure for subsequent orthogonality constraint optimization.

[0043] Taking a multi-channel advertising campaign on an e-commerce platform as an example, the platform promotes the same smartwatch product across three platforms: WeChat Moments, Douyin (TikTok) feed, and Baidu search. The conversion behavior patterns include three typical paths: "browse-compare-purchase," "impulsive click-order immediately," and "favorite-multiple visits-purchase." When constructing the adjacency matrix of the conversion behavior patterns, a 3×3 matrix is ​​used to represent the association between the three behavior paths. The connection weight between the "browse-compare-purchase" and "favorite-multiple visits-purchase" paths is 0.73, indicating a higher probability of conversion between these two behavior patterns. Based on the adjacency matrix, two main connectivity components are calculated: the first component contains rational decision-making path nodes, and the second component covers impulsive consumption path nodes. The dimensional mapping process projects the original 128-dimensional user behavior feature vector onto 64-dimensional and 32-dimensional feature subsets, corresponding to different topological substructures. The variance contribution rate calculation shows that the contribution rate of the rational decision-making feature subset is 0.67, and the variance contribution rate of the impulsive consumption feature subset is 0.31, both exceeding the preset threshold of 0.25. Therefore, both feature subsets are determined as effective segmentation units. Based on the effective segmentation units, the users are divided into 5 initial behavioral cluster sets, each containing 200-500 users with similar conversion path preferences.

[0044] The feature segmentation method guided by path topology improved the classification accuracy of user behavior features by 23.4%, and the cohesion coefficient of the initial behavior cluster set reached 0.82, significantly better than the baseline method's 0.64. This method effectively solves the curse of dimensionality problem in high-dimensional user behavior feature space and reduces the computational complexity of orthogonal projection algorithms.

[0045] Please see Figure 2 , Figure 2 This is a schematic block diagram of an AI deep analysis device for multi-platform advertising performance data provided in an embodiment of this application. The AI ​​deep analysis device 200 for multi-platform advertising performance data is used to execute the aforementioned AI deep analysis method for multi-platform advertising performance data. The AI ​​deep analysis device 200 for multi-platform advertising performance data can be configured in a server.

[0046] The server can be a standalone server, a server cluster, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0047] like Figure 2As shown, the AI ​​deep analysis device 200 for multi-platform advertising performance data includes: a graph generation module 201, a behavior analysis module 202, a data segmentation module 203, a scoring analysis module, and a scoring analysis module 204.

[0048] The graph generation module 201 is used to incrementally update the advertising performance data of each advertising platform based on a preset spatiotemporal decay factor, and generate a conversion behavior graph.

[0049] The behavior analysis module 202 is used to perform correlation analysis on the dynamic conversion behavior map to obtain conversion behavior patterns.

[0050] The data segmentation module 203 is used to divide advertising performance data into associated data blocks according to conversion behavior patterns. Each associated data block corresponds to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different associated data blocks satisfy orthogonality constraints.

[0051] The scoring and analysis module 204 is used to input related data blocks with the same conversion behavior pattern into a preset analysis model to obtain the delivery contribution score of each delivery platform.

[0052] This application provides a server, which includes a memory and a processor; the memory is used to store computer programs; the processor is used to execute the computer programs and, when executing the computer programs, implements an AI deep analysis method for multi-platform advertising performance data as described in any of the embodiments of this application.

[0053] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it enables the processor to implement an AI-based deep analysis method for multi-platform advertising performance data as described in any of the embodiments of this application.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based deep analysis method for multi-platform advertising performance data, characterized in that, The method includes: Based on a preset spatiotemporal decay factor, the advertising performance data of each advertising platform is incrementally updated to generate a conversion behavior map. Correlation analysis was performed on the dynamic transformation behavior map to obtain the transformation behavior pattern; The advertising performance data is divided into associated data blocks according to the conversion behavior pattern. Each associated data block corresponds to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different associated data blocks satisfy the orthogonality constraint. By inputting associated data blocks with the same conversion behavior pattern into a preset analysis model, a delivery contribution score for each delivery platform is obtained.

2. The AI-based deep analysis method for multi-platform advertising performance data as described in claim 1, characterized in that, Before incrementally updating the advertising performance data for each platform based on a preset spatiotemporal decay factor, the method further includes: A preset spatiotemporal decay factor is generated based on user behavior timestamps and platform environment change rate. The platform environment change rate includes: platform traffic fluctuation coefficient and competitive advertising density index.

3. The AI-based deep analysis method for multi-platform advertising performance data as described in claim 1, characterized in that, The correlation analysis of the dynamic transformation behavior map to obtain the transformation behavior pattern includes: Extract the node feature vectors and edge weight information from the transformation behavior graph; Based on the node feature vectors and the edge weight information, identify key transformation paths; The key conversion paths are summarized into patterns to generate conversion behavior patterns.

4. The AI-based deep analysis method for multi-platform advertising performance data as described in claim 3, characterized in that, The step of dividing the advertising performance data into related data blocks based on the conversion behavior pattern includes: Based on the conversion behavior pattern, cluster analysis is performed on the advertising effect data to generate user behavior clusters. The similarity of the feature vectors of different user behavior clusters satisfies the preset orthogonality constraint. Assign a unique behavior pattern tag to each user behavior cluster and establish a binding relationship table between the tag and the advertising performance data; Based on the binding relationship table, the advertising performance data is divided into multiple related data blocks.

5. The AI-based deep analysis method for multi-platform advertising performance data as described in claim 1, characterized in that, The step of inputting associated data blocks with the same conversion behavior pattern into a preset analysis model to obtain a delivery contribution score includes: The associated data block is input into a pre-set parallelized graph neural network to obtain the projection feature representation; Based on the aforementioned delivery feature representation, an adversarial evaluation model is constructed to calculate the delivery contribution score for each delivery platform.

6. The AI-based deep analysis method for multi-platform advertising performance data as described in claim 4, characterized in that, The step involves clustering the advertising performance data based on the conversion behavior pattern to generate user behavior clusters. The feature vector similarity of different user behavior clusters satisfies a preset orthogonality constraint, including: Extract user behavior features from the advertising performance data; Based on the path topology of the conversion behavior pattern, the user behavior features are segmented to generate an initial set of behavior clusters; Calculate the centroid cosine similarity of different behavior clusters in the initial behavior cluster set to generate a similarity matrix; The similarity matrix is ​​adjusted by an orthogonal projection algorithm so that the centroid cosine similarity is less than or equal to a preset threshold, thereby obtaining a set of user behavior clusters.

7. The AI-based deep analysis method for multi-platform advertising performance data as described in claim 6, characterized in that, The path topology based on the conversion behavior pattern segments the user behavior features to generate an initial set of behavior clusters, including: Construct an adjacency matrix for the transformation behavior pattern, where the element values ​​of the adjacency matrix represent the connection weights between path nodes; Based on the adjacency matrix, the connectivity components of the path topology are calculated, and path nodes with the same connectivity components are classified into the same topological substructure. Based on the node distribution characteristics of the topological substructure, the user behavior characteristics are dimension-mapped to generate a feature subset corresponding to the topological substructure; Calculate the variance contribution rate of user behavior features within each feature subset, and use the feature subsets with variance contribution rates greater than a preset threshold as effective segmentation units; Based on the effective segmentation unit, the user behavior features are clustered and grouped to generate an initial set of behavior clusters.

8. An AI-powered deep analysis device for multi-platform advertising performance data, characterized in that, The AI ​​deep analysis device for multi-platform advertising performance data is used to execute the AI ​​deep analysis method for multi-platform advertising performance data as described in any one of claims 1-7, and the AI ​​deep analysis device for multi-platform advertising performance data includes: The graph generation module is used to incrementally update the advertising performance data of each advertising platform based on a preset spatiotemporal decay factor, and generate a conversion behavior graph. The behavior analysis module is used to perform correlation analysis on the dynamic conversion behavior map to obtain conversion behavior patterns; A data segmentation module is used to divide the advertising performance data into associated data blocks according to the conversion behavior pattern. Each associated data block corresponds to a unique conversion behavior pattern, and the conversion behavior patterns corresponding to different associated data blocks satisfy orthogonality constraints. The scoring and analysis module is used to input associated data blocks with the same conversion behavior pattern into a preset analysis model to obtain the delivery contribution score of each delivery platform.

Images