Portrayal marking method and device based on flow data, and computer equipment

Abstract

The invention relates to a portrait marking method and device based on flow data, and computer equipment. The method comprises the following steps: performing multi-level time sequence analysis on historical active flow data, historical interactive flow data and historical passive flow data of a target object to generate an initial marking portrait; according to the current active flow data, the current interactive flow data and the current passive flow data of the target object, updating the features and the label of the initial marking portrait to obtain a current marking portrait; the classified scene flow data, the classified field flow data and the current marking portrait of the target object are mapped into the same embedding space, and a fused marking portrait is obtained; and according to the current active flow data, the current interactive flow data and the current passive flow data, adjusting a portrait output style of the fused marking portrait to obtain a target marking portrait. According to the method, the accuracy of the image in the marking process can be improved, and potential new rules in data can be captured in time.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, device, and computer equipment for image marking based on flow data. Background Art

[0002] Portrait tagging in traditional technology is mainly used to label and classify data. The system extracts key features from the raw data, such as users' basic information, behavioral data, transaction records, etc.; these features are further analyzed through machine learning models (such as classifiers, clustering algorithms, etc.) to generate different labels to describe the multidimensional attributes or categories of the data. These labels can help companies better understand users, market trends and business needs, and are widely used in personalized recommendations, risk prediction, customer segmentation and other fields.

[0003] However, traditional technologies rely too much on historical data and static features, and have poor adaptability to dynamic changes in data and complex scenarios, resulting in a decrease in accuracy in the image labeling process and an inability to capture potential new patterns in the data in a timely manner. Summary of the invention

[0004] Based on this, it is necessary to provide a portrait marking method, device, computer equipment, computer-readable storage medium and computer program product based on flow data, which can improve the accuracy of images during the marking process and timely capture potential new rules in the data in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for labeling a portrait based on pipeline data, comprising: Perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial labeling portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object, the features and labels of the initial marking portrait are updated to obtain the current marking portrait; Mapping the flow data of each classification scene, the flow data of each classification field and the current marking portrait of the target object into the same embedding space to obtain a fused marking portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fused marking image is adjusted to obtain the target marking image.

[0006] In a second aspect, the present application also provides a portrait marking device based on flow data, comprising: The marking image generation module is used to perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial marking image; A marking image updating module, used for updating the features and labels of the initial marking image according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object to obtain a current marking image; An image data fusion module, used to map the pipeline data of each classification scene, the pipeline data of each classification field and the current marking image of the target object into the same embedding space to obtain a fused marking image; The image style optimization module is used to adjust the image output style of the fused marking image according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data to obtain a target marking image.

[0007] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented: Perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial labeling portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object, the features and labels of the initial marking portrait are updated to obtain the current marking portrait; Mapping the flow data of each classification scene, the flow data of each classification field and the current marking portrait of the target object into the same embedding space to obtain a fused marking portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fused marking image is adjusted to obtain the target marking image.

[0008] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented: Perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial labeling portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object, the features and labels of the initial marking portrait are updated to obtain the current marking portrait; Mapping the flow data of each classification scene, the flow data of each classification field and the current marking portrait of the target object into the same embedding space to obtain a fused marking portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fused marking image is adjusted to obtain the target marking image.

[0009] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor: Perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial labeling portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object, the features and labels of the initial marking portrait are updated to obtain the current marking portrait; Mapping the flow data of each classification scene, the flow data of each classification field and the current marking portrait of the target object into the same embedding space to obtain a fused marking portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fused marking image is adjusted to obtain the target marking image.

[0010] The above-mentioned portrait labeling method, device, computer equipment, storage medium and computer program product based on pipeline data generate an initial labeled portrait by performing multi-level time series analysis on the historical active pipeline data, historical interactive pipeline data and historical passive pipeline data of the target object; update the features and labels of the initial labeled portrait according to the current active pipeline data, current interactive pipeline data and current passive pipeline data of the target object to obtain the current labeled portrait; map the pipeline data of each classified scene, the pipeline data of each classified field and the current labeled portrait of the target object into the same embedding space to obtain a fused labeled portrait; adjust the portrait output style of the fused labeled portrait according to the current active pipeline data, current interactive pipeline data and current passive pipeline data to obtain the target labeled portrait.

[0011] By performing multi-level time series analysis on the historical active, interactive and passive pipeline data of the target object, a preliminary labeling portrait can be constructed. This process provides a solid data foundation for the feature extraction of the target object and ensures a comprehensive understanding of the target object's behavior pattern. On this basis, the portrait is dynamically updated in combination with the current pipeline data to ensure that the portrait can always reflect the latest behavioral characteristics and change trends of the target object, avoiding the lag that may be caused by static portraits. Furthermore, mapping various pipeline data (such as scene and domain data) of the target object and the current labeling portrait into the same embedding space can achieve the comprehensive fusion of multi-dimensional data, improve the accuracy and comprehensiveness of the portrait, and make the portrait not only stay on the characterization of a single dimension, but more comprehensively reflect the complex behavioral characteristics of the target object. In addition, by further adjusting the fused portrait and optimizing the portrait output style according to the current pipeline data, personalized portrait customization can be achieved to meet the needs of different application scenarios. The combination of this series of technical means can improve the accuracy of images during the marking process, timely capture potential new patterns in the data, and provide more powerful and flexible support for applications such as decision support and risk prediction, thereby significantly improving the value and operability of data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 This is an application environment diagram of a portrait marking method based on pipeline data in one embodiment; Figure 2 A schematic diagram of a process of a portrait marking method based on flow data in one embodiment; Figure 3 A schematic diagram of a process for obtaining a fusion marking image in one embodiment; Figure 4 It is a flowchart of a method for obtaining a cross-domain fusion vector in one embodiment; Figure 5 A schematic diagram of a flow chart of a method for obtaining a spatial embedding vector in one embodiment; Figure 6 Schematic diagram of a flow chart of a method for obtaining a nonlinear relationship vector in one embodiment; Figure 7 is a schematic flow chart of a spatial embedding vector method in another embodiment; Figure 8A schematic diagram of a process for obtaining a target marking image in one embodiment; Fig. 9 A schematic diagram of a flow chart of a method for reconstructing a marked image in one embodiment; Fig.10 It is a structural block diagram of a portrait marking device based on flow data in one embodiment; Fig.11 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0015] The embodiment of the present application provides a method for marking an image based on flow data, which can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 through a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. The server 104 obtains the historical active flow data, historical interactive flow data and historical passive flow data of the target object in the terminal 102; and obtains the current active flow data, current interactive flow data and current passive flow data of the target object at the same time; and obtains at the same time; further, the server 104 performs multi-level time series analysis on the historical active flow data, historical interactive flow data and historical passive flow data of the target object to generate an initial labeled portrait; according to the current active flow data, current interactive flow data and current passive flow data of the target object, the features and labels of the initial labeled portrait are updated to obtain the current labeled portrait; the flow data of each classified scene, the flow data of each classified field and the current labeled portrait of the target object are mapped to the same embedding space to obtain a fused labeled portrait; according to the current active flow data, current interactive flow data and current passive flow data, the portrait output style of the fused labeled portrait is adjusted to obtain the target labeled portrait. Among them, the server 104 can be implemented with an independent server or a server cluster composed of multiple servers.

[0016] In an exemplary embodiment, Figure 2 As shown in the figure, a portrait marking method based on pipeline data is provided, and the method is applied to Figure 1 The server in the example is used to illustrate, including the following steps 202 to 208. Among them: Step 202, perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial labeling portrait.

[0017] Among them, the target object can be an individual or entity that needs to be characterized and studied during the data analysis process. It can be a user, customer, device, account, etc., and usually has a series of quantifiable behavioral data and characteristics.

[0018] The historical active transaction data may be the operation or behavior data actively initiated by the target object in the past period of time, for example, the user actively clicks on an advertisement, initiates a transaction request, or submits a query.

[0019] The historical interaction flow data may be a record of interaction behaviors between the target object and other entities, such as social interactions between users and other users, and communication between users and service platforms.

[0020] Among them, historical passive flow data can be related records generated by the system or external environment automatically sending to the target object without the target object actively initiating the behavior in the past period of time, such as the user's browsing history, log data generated by the system background, etc.

[0021] Among them, multi-level time series analysis can be a multi-dimensional and multi-level in-depth analysis method for time series data. Through different time scales (such as hours, days, weeks, months) and different dimensions (such as active, interactive, passive) analysis, it can capture the dynamic changes and trends of the target object from multiple levels, providing a full range of perspectives for portrait construction.

[0022] Among them, the initial labeling portrait can be a preliminary portrait generated based on the historical data of the target object, which includes the basic behavioral characteristics and labels of the target object.

[0023] Specifically, multi-level time series analysis of the historical active flow data, historical interactive flow data, and historical passive flow data of the target object can be carried out to finely divide different types of historical data in the time dimension and capture the behavior patterns and potential laws of the target object. Through multi-dimensional data fusion, combined with data preprocessing and feature extraction technology, an initial labeled image of the target object is constructed, including its behavioral characteristics, attribute labels, and historical trends, providing a basis for subsequent analysis.

[0024] Step 204, based on the current active flow data, the current interactive flow data and the current passive flow data of the target object, the features and labels of the initial labeled image are updated to obtain the current labeled image.

[0025] Among them, the current active flow data can be the operation or behavior data actively initiated by the target object in the current period. Different from the historical active flow data, the current data reflects the active status and behavior trend of the target object at the latest time point.

[0026] Among them, the current interaction flow data can be a record of the interaction behavior between the target object and other entities in the current time period, reflecting the target object's social activity, interaction quality and connection strength with the external environment in the current time period.

[0027] Among them, the current passive flow data can be the behavior records that the target object did not actively initiate in the current period and are automatically generated by the system or external environment. It mainly reflects the potential interests or needs of the target object and is an indirect reflection of the current state of the target object.

[0028] Among them, the current marking portrait can be a portrait updated based on the initial marking portrait according to the current data of the target object, which can more accurately reflect the behavioral characteristics and status of the target object at the latest time point.

[0029] Specifically, based on the current active flow data, interactive flow data, and passive flow data of the target object, an update algorithm is used to analyze and identify the behavioral changes or trends of the target object at the current moment, and the portrait is dynamically updated in combination with a machine learning model (such as incremental learning, online learning, etc.), so that the current labeled portrait can accurately reflect the latest behavioral characteristics and status changes of the target object.

[0030] Step 206, map the target object's classification scene flow data, classification field flow data and the current marking portrait into the same embedding space to obtain a fused marking portrait.

[0031] Among them, the classified scenario flow data can be flow data related to the target object behavior collected for different business scenarios (such as marketing, customer service, transactions, etc.).

[0032] Among them, the classified field flow data can be flow data collected according to different fields involved in the target object (such as finance, health, education, etc.).

[0033] Among them, the fused labeling portrait can be the result of fusing the flow data of the target object in various classification scenarios and classification fields with the current labeling portrait.

[0034] Specifically, the pipeline data of the target object in different classification scenarios (such as marketing scenarios, service scenarios) and the pipeline data in various classification fields (such as finance, health, social, etc.) are combined with the current labeling portrait. Since data from different sources have different characteristics and structures, it is necessary to unify the data format and dimension through feature extraction and data preprocessing. Then, these heterogeneous data are mapped into a unified embedding space using high-dimensional embedding technology (such as multidimensional scaling, autoencoders or transformers in deep learning models). In this embedding space, various types of data can be comprehensively expressed in the form of vectors, eliminating the original dimensional differences and structural inconsistencies, so that the fused labeling portrait can comprehensively and accurately represent the various behavioral characteristics of the target object. This fusion process not only integrates historical and current data, but also combines the characteristics of different scenarios and fields, providing a multi-dimensional, cross-scenario full picture, thereby greatly improving the accuracy and applicability of the target portrait.

[0035] Step 208, according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fused marking image is adjusted to obtain the target marking image.

[0036] The portrait output style can be the display form or structure of the target labeled portrait, which determines how the portrait is presented to the user or system. Adjustment of the output style can optimize the readability, accuracy and operability of the portrait. The output style may vary depending on the behavioral characteristics of the target object and the application scenario.

[0037] Among them, the target labeling portrait can be the final generated portrait that comprehensively reflects the behavioral characteristics and status of the target object.

[0038] Specifically, after generating the fusion labeling portrait, the output style of the portrait needs to be adjusted according to the current pipeline data of the target object to better adapt it to different business needs and application scenarios. According to the current active, interactive and passive pipeline data, the latest trends or changes in the behavior of the target object are identified. These data can reflect real-time information on its interests, preferences, activity frequency, etc. In combination with these changes, the various feature values ​​and display forms of the fusion portrait are carefully adjusted to optimize the readability and adaptability of the portrait. In different application scenarios, the display method of the portrait will be different. For example, in risk assessment, it may focus on highlighting potential risk points, while in a personalized recommendation system, it may focus on highlighting the preferences or needs of the target object. By adjusting the output style, the target labeling portrait is more in line with specific business needs, which improves the accuracy, operability and practical value of the portrait, thereby providing stronger data support for decision-making and subsequent operations.

[0039] In the above-mentioned portrait labeling method based on pipeline data, an initial labeled portrait is generated by performing multi-level time series analysis on the historical active pipeline data, historical interactive pipeline data and historical passive pipeline data of the target object; the features and labels of the initial labeled portrait are updated according to the current active pipeline data, current interactive pipeline data and current passive pipeline data of the target object to obtain the current labeled portrait; the pipeline data of each classified scene, the pipeline data of each classified field and the current labeled portrait of the target object are mapped to the same embedding space to obtain a fused labeled portrait; the portrait output style of the fused labeled portrait is adjusted according to the current active pipeline data, current interactive pipeline data and current passive pipeline data to obtain the target labeled portrait.

[0040] By performing multi-level time series analysis on the historical active, interactive and passive pipeline data of the target object, a preliminary labeling portrait can be constructed. This process provides a solid data foundation for the feature extraction of the target object and ensures a comprehensive understanding of the target object's behavior pattern. On this basis, the portrait is dynamically updated in combination with the current pipeline data to ensure that the portrait can always reflect the latest behavioral characteristics and change trends of the target object, avoiding the lag that may be caused by static portraits. Furthermore, mapping various pipeline data (such as scene and domain data) of the target object and the current labeling portrait into the same embedding space can achieve the comprehensive fusion of multi-dimensional data, improve the accuracy and comprehensiveness of the portrait, and make the portrait not only stay on the characterization of a single dimension, but more comprehensively reflect the complex behavioral characteristics of the target object. In addition, by further adjusting the fused portrait and optimizing the portrait output style according to the current pipeline data, personalized portrait customization can be achieved to meet the needs of different application scenarios. The combination of this series of technical means can improve the accuracy of images during the marking process, timely capture potential new patterns in the data, and provide more powerful and flexible support for applications such as decision support and risk prediction, thereby significantly improving the value and operability of data analysis.

[0041] In an exemplary embodiment, Figure 3 As shown, the target object's classification scene flow data, classification field flow data and the current marking image are mapped to the same embedding space to obtain a fused marking image, including steps 302 to 306. Among them: Step 302 , cross-domain feature fusion is performed on the pipeline data of each classification scene and the pipeline data of each classification field to obtain cross-domain fusion vectors.

[0042] Among them, cross-domain feature fusion can be the integration of data features from different scenarios and fields in order to comprehensively analyze the behavioral characteristics of the target object from multiple angles.

[0043] Among them, the cross-domain fusion vector can be a high-dimensional vector representation obtained by cross-domain feature fusion, which includes the behavioral characteristics of the target object in different domains and scenarios.

[0044] Specifically, it is necessary to integrate the pipeline data from different sources (including pipeline data of each classification scene and pipeline data of classification field) into a unified representation. Each data type usually has different characteristics and structures, so it is necessary to extract features and standardize them for fusion processing. Feature alignment technology is used to convert different types of data into the same dimension and normalize the data so that they can be compared in the same space. On this basis, feature fusion is performed through weighted averaging, splicing or using deep learning models (such as multi-layer perceptrons or autoencoders), and the key and auxiliary information in the data of each scene and each field are merged into various cross-domain fusion vectors. Each vector not only contains the basic information of each type of data, but also effectively integrates the behavior patterns of the target object in different scenes and fields to form a more comprehensive behavioral feature representation.

[0045] Step 304: Identify nonlinear relationship information from each cross-domain fusion vector to obtain each spatial embedding vector.

[0046] Among them, nonlinear relationship information can be the complex dependencies between the features in the data, which cannot be captured by traditional linear models (such as linear regression, simple weighted sum, etc.).

[0047] Among them, the spatial embedding vector can be a vector representation that maps high-dimensional and complex original data into a low-dimensional continuous space through embedding learning technology.

[0048] Specifically, using nonlinear mapping and deep learning techniques, such as multi-layer neural networks or convolutional neural networks, the complex associations in the data can be captured in the high-dimensional feature space, especially the mutual influence between target objects in different scenarios and fields. At the same time, nonlinear dimensionality reduction techniques (such as t-SNE or autoencoders) can also help identify the implicit structure and association in the data, revealing the potential nonlinear relationship between the behaviors of the target objects, and extracting the nonlinear relationship information composed of these complex associations and nonlinear relationships through activation functions and weight adjustments, and finally converting them into spatial embedding vectors.

[0049] Step 306, mapping each spatial embedding vector and the current marking image into the same embedding space to obtain a fused marking image.

[0050] Specifically, features from different sources (including spatial embedding vectors extracted from cross-domain fusion vectors and existing current labeled portraits) are integrated into a comprehensive and unified vector representation, which is achieved through embedding learning methods, such as using shared weight networks or generative adversarial networks (GANs) in deep learning to map features from different data sources to the same embedding space. In this embedding space, the current labeled portrait and the spatial embedding vector share the same feature dimension, allowing them to interact and perform further analysis. The generated fused labeled portrait not only integrates the target object's behavioral information in different scenarios and fields, but also accurately reflects its latest status and dynamic changes, and ultimately obtains a portrait that can provide comprehensive support for subsequent analysis and decision-making.

[0051] In this embodiment, by performing cross-domain feature fusion of classified scene pipeline data and classified domain pipeline data, the obtained cross-domain fusion vector can effectively combine different types of data features, thereby improving the correlation and expressiveness between data. Then, by identifying the nonlinear relationship information in the cross-domain fusion vector, the generated spatial embedding vector further extracts the potential patterns and relationships in the data, and can more accurately capture complex feature interactions. Finally, these spatial embedding vectors are mapped to the same embedding space with the current marked portrait, and the generated fused marked portrait is not only highly personalized and customized visually, but also can more accurately reflect the association between data features and user behavior. This process can significantly improve the adaptability, accuracy and information communication efficiency of the portrait, ultimately optimizing the user experience and enhancing the effect of data-driven decision-making.

[0052] In an exemplary embodiment, Figure 4 As shown, cross-domain feature fusion is performed on the pipeline data of each classification scene and the pipeline data of each classification field to obtain each cross-domain fusion vector, including steps 402 to 406. Among them: Step 402 , extracting features from the pipeline data of each classification scene and the pipeline data of each classification field to obtain the pipeline feature vectors of each classification scene and the pipeline feature vectors of each classification field.

[0053] The classification scene pipeline feature vector may be a vector representation obtained after feature extraction of pipeline data of a target object in a specific scene.

[0054] The classification domain pipeline feature vector may be a vector representation obtained after feature extraction of pipeline data of a target object in a specific domain.

[0055] Specifically, in the feature extraction stage, the pipeline data of each classification scene and each classification field are first preprocessed to remove noise and standardize. Then, important features are extracted from the original data through statistical analysis, time series modeling or deep learning methods (such as CNN, LSTM, etc.). For the classification scene pipeline data, basic statistical features (such as mean, variance, etc.) can be extracted, and for data with time series properties, autoregressive models or LSTM are used to capture the temporal patterns in the data. In addition, the use of deep learning models (such as CNN) can automatically extract high-order features from the original data, and finally generate feature vectors for each data set, including classification scene pipeline feature vectors and classification field pipeline feature vectors.

[0056] Step 404, traverse the mutual relationship between the flow feature vectors of each classification scene and the flow feature vectors of each classification field to obtain a classification feature vector relationship network.

[0057] Among them, the classification feature vector relationship network can be a network structure constructed based on the relationship between each classification scene pipeline feature vector and each classification field pipeline feature vector, which includes the direct closeness information between each classification scene pipeline feature vector and each classification field pipeline feature vector.

[0058] Specifically, when constructing the classification feature vector relationship network, the similarity between the classification scene pipeline feature vector and the classification field pipeline feature vector is calculated, using metrics such as cosine similarity, Pearson correlation coefficient, or Euclidean distance. A graph network is constructed through similarity, in which each feature vector represents a node, and the similarity between features determines the weight of the edge. The graph network can help identify the dependencies and influences between different features. In addition, the use of graph neural networks (GNNs) can further identify the complex relationships between graph network feature vectors, thereby obtaining a more accurate classification feature vector relationship network.

[0059] Step 406 , mutually migrate the feature information of the flow feature vectors of each classification scene and the flow feature vectors of each classification domain whose vector relationship closeness value in the classification feature vector relationship network is greater than a preset closeness threshold, to obtain each cross-domain fusion vector.

[0060] The vector closeness value may be a value measuring the similarity or correlation between two vectors.

[0061] The preset close threshold may be used to determine whether two vectors have sufficient similarity or relationship.

[0062] Specifically, based on the classification feature vector relationship network, the vector relationship closeness values ​​between the pipeline feature vectors of each classification scene and the pipeline feature vectors of each classification field are determined. Transfer learning or self-attention mechanism is used to share the knowledge between the classification scene pipeline feature vectors and the classification field pipeline feature vectors whose vector relationship closeness values ​​are greater than the preset closeness threshold, so that the data of the two fields can influence and complement each other. Through the self-attention mechanism, the weights of different features are automatically adjusted to dynamically select and fuse the most important feature information. Finally, feature mapping (such as a fully connected layer or a convolutional layer) is used to fuse the two feature vectors whose closeness is greater than the preset closeness threshold into a cross-domain fusion vector. This vector contains rich information from the classification scene and the classification field, providing a comprehensive feature representation for subsequent analysis tasks.

[0063] In this embodiment, by extracting features from the classification scene pipeline data and the classification field pipeline data, the obtained classification scene pipeline feature vectors and classification field pipeline feature vectors provide a multi-dimensional representation for the data, which can fully capture the key features of different categories and fields. Further, by traversing the mutual relationships between the feature vectors, the constructed classification feature vector relationship network enables the correlation between different data sources to be revealed, enhancing the semantic understanding between the features. On this basis, information migration is performed for features whose vector relationship closeness exceeds a preset threshold, effectively integrating the key information of the classification scene pipeline feature vectors and the classification field pipeline feature vectors, thereby generating a cross-domain fusion vector. This process can eliminate differences between fields, improve the overall consistency and expression ability of the data, and thus provide more accurate and comprehensive feature information for subsequent analysis, prediction, and decision-making.

[0064] In an exemplary embodiment, Figure 5 As shown, nonlinear relationship information is identified from each cross-domain fusion vector to obtain each spatial embedding vector, including steps 502 to 506. Among them: Step 502: Identify nonlinear relationship information from each cross-domain fusion vector to obtain each nonlinear relationship vector.

[0065] Among them, nonlinear relationship information can be the complex interactions existing in the data, which cannot be represented by a simple linear model.

[0066] The nonlinear relationship vector may be a representation obtained by performing a nonlinear transformation on the original data features.

[0067] Specifically, since the cross-domain fusion vector may contain complex nonlinear interactions, activation functions (such as ReLU, Sigmoid, tanh, etc.) are used to perform nonlinear transformations on each cross-domain fusion vector. These activation functions can compress and expand the data to help capture complex patterns and features. In order to extract nonlinear relationships from complex data, deep neural networks (DNNs), especially multi-layer perceptrons (MLPs), can also be used to automatically learn the nonlinear dependencies between input features through their multi-layer structures; MLPs can extract complex nonlinear features from the original input data through layer-by-layer weight matrix transformations. After these processes, each cross-domain fusion vector will be mapped to a new space, and the generated nonlinear relationship vector can effectively express the nonlinear features in the data.

[0068] Step 504: Perform high-order feature interactions on each nonlinear relationship vector to obtain each interaction embedding vector.

[0069] Among them, high-order feature interactions can be complex combinations of multiple features. By combining and interacting originally simple low-order features, deeper dependencies between them can be captured.

[0070] Among them, the interactive embedding vector can be a new feature vector obtained after high-order feature interaction between multiple features.

[0071] Specifically, feature interaction layers (such as multi-layer perceptrons, deep neural networks, etc.) are introduced to capture the complex nonlinear interactions between different features. In high-order feature interactions, each pair of nonlinear relationship vectors will interact with other features through combinations (such as weighted summation or concatenation) to form new and more expressive feature representations. The results of these interactions will produce interactive embedding vectors that can better represent the complex relationships between the original features and provide richer information for subsequent ones.

[0072] Step 506: Adjust the adaptation range of each interactive embedding vector to obtain each spatial embedding vector.

[0073] Specifically, after the high-order features interact, regularization and constraint adjustment methods are needed to adjust the adaptability of the interaction embedding vector so that it can have better generalization capabilities in different tasks or data scenarios. Through these adjustments, the interaction embedding vector will be optimized into spatial embedding vectors that can adapt to a wider range of scenarios and show richer and more accurate data representation.

[0074] In this embodiment, by identifying nonlinear relationship information from the cross-domain fusion vector, more complex and deep patterns in the data can be captured, which goes beyond the limitations of linear relationships and improves the ability to represent features. Then, by performing high-order feature interactions on the nonlinear relationship vector, the generated interactive embedding vector effectively integrates the high-order dependencies between multiple features, further enhancing the ability to understand complex data structures. On this basis, by adjusting the adaptive range of the interactive embedding vector, the generated spatial embedding vector can adaptively optimize the representation of the feature space, so that the final vector can more accurately adapt to different data scenarios and requirements, thereby improving the generalization ability and expressiveness of the data in practical applications. Through this series of processing, the accuracy of the features can be greatly improved, providing more powerful support for data analysis and decision-making.

[0075] In an exemplary embodiment, Figure 6 As shown, nonlinear relationship information is identified from each cross-domain fusion vector to obtain each nonlinear relationship vector, including steps 602 to 604. Among them: Step 602 , performing high-order transformation on each cross-domain fusion vector to obtain each cross-domain high-order vector.

[0076] Among them, high-order transformation can be to expand and transform the original feature vector through mathematical operations and feature combination methods to generate a more complex feature representation.

[0077] Among them, cross-domain high-order vectors can be vector representations obtained by performing high-order transformations on data features. These vectors not only contain the basic features in the original data, but also integrate multiple high-order combination relationships of cross-domain data.

[0078] Specifically, polynomial feature expansion or neural network hierarchy (such as fully connected layer) is used to perform high-order mapping on each cross-domain fusion vector, and the new cross-domain high-order vector will contain more feature combinations and better represent the complex patterns in the data. For example, for two original feature vectors A and B, high-order transformation can obtain richer feature expressions by calculating the product of A and B (A×B) or other high-order combinations.

[0079] Step 604: Perform nonlinear transformation on each cross-domain high-order vector to obtain each nonlinear relationship vector.

[0080] Among them, nonlinear transformation can be to apply nonlinear functions (such as ReLU, Sigmoid, tanh, etc.) to transform the feature vector so that the nonlinear relationship in the data can be revealed and captured.

[0081] Specifically, activation functions (such as ReLU, Sigmoid, tanh, etc.) are usually used to perform nonlinear transformations on cross-domain high-order vectors, where the activation function can change the distribution of the feature space so that the model can output more complex, nonlinear feature interactions during the calculation process. For example, using the ReLU activation function will "suppress" the negative value part to zero, allowing the model to focus on capturing the positive relationship of the feature. After nonlinear transformation, the original cross-domain high-order vectors will be converted into nonlinear relationship vectors, which can effectively express the complex interactive relationships in cross-domain data and provide rich information for subsequent learning and prediction tasks.

[0082] In this embodiment, by performing high-order transformation on the cross-domain fusion vector, more complex feature relationships can be mined, the feature space is expanded, and the depth and breadth of data expression are improved. This high-order transformation can capture the potential high-order dependencies between features and enhance the understanding and learning ability of complex data patterns. Subsequently, by performing nonlinear transformation on the cross-domain high-order vector, the nonlinear relationship in the data can be further mined, thereby breaking through the limitations of the linear model and obtaining a more accurate feature representation. This nonlinear transformation can identify and learn more complex patterns and associations, and improve the expressiveness of the data. This processing process helps to better understand and utilize the deep features in cross-domain data, and enhances the performance and generalization capabilities in practical applications.

[0083] In an exemplary embodiment, Figure 7 As shown, the adaptation range of each interactive embedding vector is adjusted to obtain each spatial embedding vector, including steps 702 to 710. Among them: Step 702: Perform ridge regression processing on each interactive embedding vector to obtain each regression regularized vector.

[0084] The regression regularization vector may be a feature vector processed by a regression regularization method (such as ridge regression).

[0085] Specifically, each interaction embedding vector will be used as input for ridge regression, where ridge regression is a linear regression method that penalizes large parameter values ​​in the model by adding an L2 regularization term to the loss function, thereby reducing the risk of overfitting. Specifically, the interaction embedding vector is input into the ridge regression model, which adjusts the weight of the vector by minimizing the loss function with L2 regularization, so that the objective function can fit the data with small parameter changes, thereby obtaining a regression regularized vector, that is, the result of regularizing the original interaction embedding vector.

[0086] Step 704, regularize the domain of each regression regularization vector to obtain each domain regularization vector.

[0087] The domain regularization vector may be a vector that further optimizes and adjusts the feature vector after regression regularization to adapt to feature distribution in different domains.

[0088] Specifically, each regression regularization vector after ridge regression regularization is subjected to domain regularization processing, where the goal of domain regularization is to adjust each regression regularization vector according to the characteristics of different fields, so that the features are more adaptable between fields. It usually includes the introduction of domain-specific regularization terms, which can be achieved by designing domain-specific loss functions or regularization strategies. For example, domain adaptation technology is used or domain-related weights are added to the loss function to constrain the feature vectors of each field. These regularization strategies adjust the feature representation according to the differences between fields, so that the final feature vector can reflect the special needs of the field, and the obtained vector is the regularization vector of each field.

[0089] Step 706, regularize the scene of each domain regularization vector to obtain a regularization vector for each scene.

[0090] The scene regularization vector can be a vector that optimizes the feature vector after domain regularization to adapt to a specific application scenario. The purpose of scene regularization is to adjust the feature vector so that it has better expression ability and discrimination in a specific scenario.

[0091] Specifically, the regularized vectors of each field are further regularized for scenes, where the purpose of scene regularization is to ensure that the feature vectors have good adaptability and consistency in different application scenarios, which is achieved by introducing scene-specific regularization strategies. For example, constraints on scene-specific features or data distribution can be imposed by adding penalty terms for different scenarios to the loss function, or by applying scene adaptation methods for optimization. The goal of scene regularization is to make the feature vectors have better expressive power in different scenarios, to more accurately reflect the feature differences and similarities between scenes, and finally to obtain regularized vectors for each scene, which can better characterize the feature information in a specific scene.

[0092] Step 708: Use triplet loss to constrain the semantic consistency of each scene regularization vector to obtain each semantically constrained regularization vector.

[0093] The semantically constrained regularization vector may be a feature vector obtained after applying the triplet loss, which aims to ensure that semantically similar samples are closer in the feature space, while semantically different samples are farther away.

[0094] Specifically, a scene regularization vector is selected from each scene regularization vector as an anchor point, which represents the characteristics of the current scene. The selection of the anchor point is usually based on the needs of the current task or randomly selected from each scene regularization vector. Then, a positive sample with similar semantics to the anchor point and a negative sample with a large semantic difference from the anchor point are selected from each scene regularization vector. Through the optimization of the triplet loss, the goal is to minimize the distance between the anchor scene regularization vector and the positive sample scene vector, while ensuring that the distance between the anchor scene regularization vector and the negative sample scene vector is as large as possible, that is, the scene regularization vectors with similar semantics are closer, and the scene regularization vectors with different semantics are farther away, thereby obtaining a semantically constrained regularization vector after semantic consistency constraint.

[0095] Step 710, according to the actual application target information, reinforcement learning constraints are performed on each semantic constraint regularization vector to obtain each spatial embedding vector.

[0096] The actual application target information may be an optimization target set according to specific tasks and requirements during the calculation process.

[0097] Specifically, according to the actual application target information, the semantic constraint regularization vector is subjected to reinforcement learning constraints, where the core of reinforcement learning is to optimize the strategy through interaction with the environment to maximize the predetermined reward function. In the application of this step, reinforcement learning guides the adjustment of the feature vector by designing a reward function, which is usually closely related to the actual application goal (such as classification accuracy, recommendation effect, etc.). In specific implementation, the semantic constraint regularization vector can be adjusted using reinforcement learning algorithms (such as Q-learning, policy gradient, etc.) to improve its effect in practical applications. The vector optimized by reinforcement learning can better adapt to the needs of the task and finally obtain the spatial embedding vector.

[0098] In this embodiment, by performing ridge regression processing on the interactive embedding vector, the features can be effectively regularized, the risk of overfitting can be reduced, and better generalization to unseen data can be ensured. Then, by gradually regularizing the domains and scenarios of the regression regularization vector, the expressiveness and robustness of the features are further optimized, ensuring adaptability and consistency in different domains and scenarios. In this process, by using the triple loss to constrain the semantic consistency of the scene regularization vector, the feature semantics in different scenarios are ensured to remain consistent, avoiding information loss or incorrect association. Finally, by performing reinforcement learning constraints on the semantic constraint regularization vector, the feature vector is further adjusted and optimized according to the actual application goals to ensure that the final spatial embedding vector can accurately reflect the data characteristics and achieve higher prediction accuracy and decision-making efficiency in practical applications.

[0099] In an exemplary embodiment, Figure 8 As shown, according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fusion marking image is adjusted to obtain the target marking image, including steps 802 to 806. Among them: Step 802, setting initial dynamic customized style information according to current active flow data, current interactive flow data and current passive flow data.

[0100] The initial dynamic customized style information may be setting information of a customized marking image style obtained through analysis of data from different sources.

[0101] Specifically, key features are extracted from the input active flow data (such as user active behaviors, such as clicks, searches, purchases, etc.), interactive flow data (such as interaction data between users and the system, such as browsing time, click frequency, messages, etc.) and passive flow data (such as user geographic location, device type, network environment, etc.), and by analyzing the trends, frequencies, behavior patterns and other information of the key feature data, initial dynamic customized style information related to user behaviors and data patterns is generated.

[0102] Step 804, reconstructing the image style of the fused marking image according to the initial dynamic customized style information to obtain a reconstructed marking image.

[0103] The reconstructed marking image may be a new marking image obtained by adjustment and transformation after applying the initial dynamic customized style information.

[0104] Specifically, the original fused labeled image is reconstructed based on the extracted initial dynamic customized style information. In this process, the style of the fused labeled image (such as layout, color, font, graphic mark, etc.) will be adjusted in a targeted manner according to the characteristics of the current active flow data, the current interactive flow data, and the current passive flow data. For example, if a user's behavior is frequent, its characteristics may be highlighted by deepening the color, enlarging the icon, etc., otherwise the conspicuousness of the relevant logo may be reduced. In addition, depending on the differences in data sources, the presentation method of the image may be adapted to ensure that the image can accurately convey information in different situations and obtain a reconstructed labeled image.

[0105] Step 806, performing image perception feature optimization on the reconstructed marking image to obtain a target marking portrait.

[0106] Among them, image perception feature optimization can be achieved through a series of visual optimization techniques to improve the image's effect at the user's perception level. This process includes adjusting the image's contrast, brightness, color balance, element spacing, detail clarity, etc., to make the image more visually coordinated, clear, and easy to understand.

[0107] Specifically, the reconstructed marking image is further optimized in the direction of image perception features. The optimization process improves the visual effect of the image through image perception algorithms (such as optimization based on visual perception models), including adjusting various visual elements of the image, such as contrast, color balance, graphic spacing, element hierarchy, edge clarity, etc., to make it more in line with the perception habits of the human eye, thereby improving the clarity of the image and the efficiency of information transmission. In addition, the optimization process also considers the display effect of users on different devices, such as responsive design, to ensure that the image can maintain a good visual experience at various screen sizes and resolutions, and finally obtain a target marking portrait that meets user needs.

[0108] In this embodiment, by combining the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the set initial dynamic customized style information can generate personalized image style information according to the characteristics of different data sources. The application of this information provides precise guidance for the reconstruction of the fusion marking portrait, so that the image style can better adapt to actual needs and scene changes. Furthermore, by optimizing the image perception features of the reconstructed marking image, the visual expression and information communication effect of the image can be improved, ensuring the clarity and consistency of the image in all details. The final target marking portrait can meet the requirements of practical applications in terms of accuracy and adaptability, help improve user experience, optimize decision support, and improve the efficiency and accuracy of data-driven analysis.

[0109] In an exemplary embodiment, Fig. 9 As shown, according to the initial dynamic customized style information, the image style of the fused marking image is reconstructed to obtain a reconstructed marking image, including steps 902 to 910. Among them: Step 902, based on the initial dynamic customized style information, preliminarily reconstruct the image style of the fused marked portrait to obtain an initial reconstructed image.

[0110] Among them, the initial reconstructed image can be the first labeled image style reconstruction of the fused labeled portrait according to the set initial dynamic customized style information, which provides a basic visual expression framework and reflects the preliminary mapping of data features.

[0111] Specifically, the style parameters obtained by analyzing the changes in user information (such as active, interactive and passive flow data) are extracted from the initial dynamic customized style information. These parameters may include color, font, icon, layout structure, information display level, etc. According to these preliminary style parameters, the visual style of the original fusion marking portrait is adjusted through design rules or algorithms. For example, the size of the interactive elements may be increased or the color may be changed to improve the visualization effect according to the user's interaction frequency, or the layout of the image may be adjusted according to the user's activity status, and the initial reconstructed image is output, which visually reflects the results of the preliminary data features and style adjustments, and provides a basis for subsequent optimization and fine-tuning.

[0112] Step 904 , reverse reasoning is performed on the original style information according to the initial reconstructed image to obtain reverse dynamic customized style information.

[0113] The reverse dynamic customized style information may be style information obtained by reverse reasoning the initial reconstructed image.

[0114] Specifically, the actual style of the initial reconstructed image is analyzed, and the relationship between the original image style information and the reconstructed style is extracted through reverse reasoning methods, including quantitative analysis of each element in the initial reconstructed image (such as color matching, layout structure, font size and position, icon arrangement, etc.), from which the design intention and data-driven model behind the style are inferred. Reverse reasoning can reveal the relationship between elements in image design and how these visual adjustments respond to input data features, and obtain reverse dynamic customized style information, which indicates which visual parameters need to be further optimized or adjusted after the initial style adjustment.

[0115] Step 906 , determining style information difference data according to the difference between the reverse dynamic customized style information and the initial dynamic customized style information.

[0116] Among them, the style information difference data can be a quantitative result of the difference between the reverse dynamic customized style information and the initial dynamic customized style information, which clarifies the specific direction and gap of the style adjustment.

[0117] Specifically, by comparing the specific differences between the reverse dynamic customized style information and the initial dynamic customized style information, the system can identify changes and deviations that occur during the image reconstruction process, such as which visual elements such as color, icon size, layout position, etc. are different. The sources of these differences may include errors in changes in data sources, errors in data on changes in user behavior patterns, or errors in the adjustment of the style design itself. This difference data helps to correct errors in style changes and provides specific target information for further optimization of the style. The final style information difference data provides clear guidance and adjustment basis for subsequent customized style optimization.

[0118] Step 908, setting and optimizing the dynamic customized style information according to the style information difference data, the current active flow data, the current interactive flow data and the current passive flow data.

[0119] Among them, optimizing the dynamic customized style information can be an optimization solution generated by combining the style information difference data with the current data flow information (such as active, interactive and passive flow data), aiming to customize and adjust the image style more finely to ensure that the image can better adapt to different application scenarios and needs.

[0120] Specifically, the image style is set more finely by combining the style information difference data with the current active flow data, the current interactive flow data, and the current passive flow data. The different difference information in the style information difference data is analyzed, and the style setting algorithm is adjusted by combining the logic and method of setting the style with the style information difference data, the current active flow data, the current interactive flow data, and the current passive flow data, so that the re-output customized data can better express the characteristics of the current data, generate optimized dynamic customized style information, ensure that the image better matches the user's needs and data characteristics in visual design, and enhance the user experience.

[0121] Step 910, based on the optimized dynamically customized style information, reconstruct the image style of the fused marking image to obtain a reconstructed marking image.

[0122] Specifically, the original fused labeled image is reconstructed using the generated optimized dynamic customized style information. The goal of the secondary reconstruction is to eliminate the style errors found in the previous stage and apply the optimization requirements to the visual design of the image to ensure that the image can accurately reflect the changes in the data and can effectively convey information visually. Specific operations include further optimizing the color matching, element layout, font size, and hierarchy of identifiers of the image to improve the visibility and readability of the image. The image after secondary reconstruction will more accurately meet the target requirements, and finally obtain a reconstructed labeled image, which not only has a visual optimization effect, but also can accurately show the characteristics and laws of the data to meet the needs of practical applications.

[0123] In this embodiment, by preliminarily reconstructing the image style of the fused marking portrait according to the initial dynamic customized style information, it is possible to find a balance between personalized needs and data features and generate a basic image style framework. Then, by reverse reasoning the initial reconstructed image, the reverse dynamic customized style information is obtained, and the potential laws of the image style and the data characteristics reflected are further explored, thereby providing more references for style optimization. Based on the difference between the reverse dynamic customized style information and the initial style information, the difference data in the style can be accurately identified, so as to make fine adjustments to the image style. On this basis, combined with the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the optimized dynamic customized style information is set, and by further optimizing the style, it is ensured that the generated reconstructed marking image more accurately reflects the needs and scene changes of the actual application. The whole process helps to generate highly customized marking images that meet actual needs, improve the accuracy, adaptability and visual effects of image expression, and optimize the performance of the decision support system.

[0124] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0125] Based on the same inventive concept, the embodiment of the present application also provides a kind of image marking device based on running data for realizing the image marking method based on running data involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of one or more image marking devices based on running data provided below can refer to the limitations of an image marking method based on running data above, and will not be repeated here.

[0126] In an exemplary embodiment, Fig.10 As shown, a portrait marking device based on pipeline data is provided, including: a marking image generation module 1002, a marking image update module 1004, an image data fusion module 1006 and an image style optimization module 1008, wherein: The marking image generation module 1002 is used to perform multi-level time series analysis on the historical active flow data, historical interactive flow data and historical passive flow data of the target object to generate an initial marking image; The marking image updating module 1004 is used to update the features and labels of the initial marking image according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object to obtain the current marking image; The image data fusion module 1006 is used to map the flow data of each classification scene, the flow data of each classification field and the current marking image of the target object into the same embedding space to obtain a fused marking image; The image style optimization module 1008 is used to adjust the image output style of the fused marking image according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data to obtain the target marking image.

[0127] In one embodiment, the image data fusion module 1006 is also used to perform cross-domain feature fusion on the pipeline data of each classified scene and the pipeline data of each classified field to obtain each cross-domain fusion vector; identify nonlinear relationship information from each cross-domain fusion vector to obtain each spatial embedding vector; map each spatial embedding vector and the current marked image to the same embedding space to obtain a fused marked image.

[0128] In one embodiment, the image data fusion module 1006 is also used to extract features from the flow data of each classification scene and the flow data of each classification field to obtain the flow feature vectors of each classification scene and the flow feature vectors of each classification field; traverse the relationship between the flow feature vectors of each classification scene and the flow feature vectors of each classification field to obtain a classification feature vector relationship network; and mutually migrate the feature information of the flow feature vectors of each classification scene and the flow feature vectors of each classification field whose vector relationship closeness value in the classification feature vector relationship network is greater than a preset closeness threshold to obtain each cross-domain fusion vector.

[0129] In one embodiment, the image data fusion module 1006 is further used to identify nonlinear relationship information from each cross-domain fusion vector to obtain each nonlinear relationship vector; perform high-order feature interaction between each nonlinear relationship vector to obtain each interactive embedding vector; and adjust the adaptation range of each interactive embedding vector to obtain each spatial embedding vector.

[0130] In one embodiment, the image data fusion module 1006 is further used to perform high-order transformation on each cross-domain fusion vector to obtain each cross-domain high-order vector; and perform nonlinear transformation on each cross-domain high-order vector to obtain each nonlinear relationship vector.

[0131] In one embodiment, the image data fusion module 1006 is also used to perform ridge regression processing on each interactive embedding vector to obtain each regression regularized vector; regularize the domain of each regression regularized vector to obtain each domain regularized vector; regularize the scene of each domain regularized vector to obtain each scene regularized vector; use triplet loss to constrain the semantic consistency of each scene regularized vector to obtain each semantic constraint regularized vector; according to actual application target information, perform reinforcement learning constraints on each semantic constraint regularized vector to obtain each spatial embedding vector.

[0132] In one embodiment, the image style optimization module 1008 is also used to set initial dynamic customized style information based on the current active pipeline data, the current interactive pipeline data and the current passive pipeline data; reconstruct the image style of the fused marking portrait according to the initial dynamic customized style information to obtain a reconstructed marking image; and optimize the image perception features of the reconstructed marking image to obtain a target marking portrait.

[0133] In one embodiment, the image style optimization module 1008 is also used to preliminarily reconstruct the image style of the fused marked portrait based on the initial dynamic customized style information to obtain an initial reconstructed image; reversely infer the original style information based on the initial reconstructed image to obtain reverse dynamic customized style information; determine the style information difference data based on the difference between the reverse dynamic customized style information and the initial dynamic customized style information; set the optimized dynamic customized style information based on the style information difference data, the current active pipeline data, the current interactive pipeline data and the current passive pipeline data; and reconstruct the image style of the fused marked portrait based on the optimized dynamic customized style information to obtain a reconstructed marked image.

[0134] Each module in the above-mentioned image marking device based on flow data can be implemented in whole or in part by software, hardware and their combination. Each of the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0135] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.11As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store server data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a portrait marking method based on pipeline data is implemented.

[0136] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0137] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0138] In one embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0139] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-mentioned method embodiments.

[0140] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0142] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A portrait marking method based on flow data, characterized in that: The method comprises: Perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial labeling portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object, the features and labels of the initial marking portrait are updated to obtain the current marking portrait; Mapping the flow data of each classification scene, the flow data of each classification field and the current marking portrait of the target object into the same embedding space to obtain a fused marking portrait; According to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data, the image output style of the fused marking image is adjusted to obtain the target marking image.

2. The method according to claim 1, characterized in that The step of mapping the classification scene flow data, classification field flow data and the current marking image of the target object into the same embedding space to obtain a fused marking image includes: Perform cross-domain feature fusion on the pipeline data of each classification scene and the pipeline data of each classification field to obtain each cross-domain fusion vector; Identifying nonlinear relationship information from each of the cross-domain fusion vectors to obtain each spatial embedding vector; Each of the spatial embedding vectors and the current marking image are mapped into the same embedding space to obtain the fused marking image.

3. The method according to claim 2, characterized in that The cross-domain feature fusion of each of the classification scene pipeline data and each of the classification field pipeline data to obtain each cross-domain fusion vector includes: Perform feature extraction on the pipeline data of each classification scene and the pipeline data of each classification field to obtain a pipeline feature vector of each classification scene and a pipeline feature vector of each classification field; Traversing the mutual relationship between the flow feature vectors of each classification scene and the flow feature vectors of each classification field to obtain a classification feature vector relationship network; The feature information of each of the classification scene flow feature vectors and each of the classification field flow feature vectors whose vector relationship closeness value in the classification feature vector relationship network is greater than a preset closeness threshold is mutually migrated to obtain each of the cross-field fusion vectors.

4. The method according to claim 2, characterized in that: The step of identifying nonlinear relationship information from each of the cross-domain fusion vectors to obtain each spatial embedding vector includes: Identifying nonlinear relationship information from each of the cross-domain fusion vectors to obtain each nonlinear relationship vector; Performing high-order feature interactions on the nonlinear relationship vectors to obtain interaction embedding vectors; The adaptation range of each of the interactive embedding vectors is adjusted to obtain each of the spatial embedding vectors.

5. The method according to claim 4, characterized in that The step of identifying nonlinear relationship information from each of the cross-domain fusion vectors to obtain each nonlinear relationship vector includes: Performing high-order transformation on each of the cross-domain fusion vectors to obtain each cross-domain high-order vector; Nonlinear transformation is performed on each of the cross-domain high-order vectors to obtain each of the nonlinear relationship vectors.

6. The method according to claim 4, characterized in that The step of adjusting the adaptation range of each of the interactive embedding vectors to obtain each of the spatial embedding vectors includes: Performing ridge regression processing on each of the interactive embedding vectors to obtain each regression regularization vector; Regularizing the domain of each regression regularization vector to obtain a regularization vector for each domain; Regularizing the scenes of the regularized vectors of the fields to obtain regularized vectors of the scenes; Using a triplet loss to constrain the semantic consistency of each of the scene regularization vectors, to obtain each of the semantically constrained regularization vectors; According to the actual application target information, each of the semantic constraint regularization vectors is subjected to reinforcement learning constraints to obtain each of the spatial embedding vectors.

7. The method according to claim 1, characterized in that The step of adjusting the image output style of the fused marking image according to the current active pipeline data, the current interactive pipeline data, and the current passive pipeline data to obtain the target marking image includes: Setting initial dynamic customized style information according to the current active flow data, the current interactive flow data and the current passive flow data; Reconstructing the image style of the fused marking image according to the initial dynamic customized style information to obtain a reconstructed marking image; The image perception features of the reconstructed marking image are optimized to obtain the target marking portrait.

8. The method according to claim 7, characterized in that The step of reconstructing the image style of the fused marking image according to the initial dynamic customized style information to obtain a reconstructed marking image includes: Preliminarily reconstructing the image style of the fused marked image according to the initial dynamically customized style information to obtain an initial reconstructed image; According to the initial reconstructed image, reverse reasoning is performed on the original style information to obtain reverse dynamic customized style information; Determining style information difference data according to a difference between the reverse dynamically customized style information and the initial dynamically customized style information; According to the style information difference data, the current active flow data, the current interactive flow data and the current passive flow data, setting and optimizing the dynamic customized style information; According to the optimized dynamically customized style information, the image style of the fused marking image is reconstructed to obtain the reconstructed marking image.

9. A portrait marking device based on flow data, characterized in that: The device comprises: The marking image generation module is used to perform multi-level time series analysis on the historical active flow data, historical interactive flow data, and historical passive flow data of the target object to generate an initial marking image; A marking image updating module, used for updating the features and labels of the initial marking image according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data of the target object to obtain a current marking image; An image data fusion module, used to map the pipeline data of each classification scene, the pipeline data of each classification field and the current marking image of the target object into the same embedding space to obtain a fused marking image; The image style optimization module is used to adjust the image output style of the fused marking image according to the current active pipeline data, the current interactive pipeline data and the current passive pipeline data to obtain a target marking image.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images