Social intelligent agent cluster control method and device

Abstract

The invention discloses a social intelligent agent device and a cluster control method and device thereof, and the method comprises the steps: carrying out the statistics of different groups of Internet platform users, and constructing corresponding user feature portraits; according to the internet user feature portraits, user portraits are formed for the intelligent agents deployed to the cluster, and the intelligent agents with preferences are obtained; creating an information consumption activity task for the agent based on a BS architecture; information consumption data resulting from agent activity is collected. According to the invention, through agent portrait formation and task scheduling optimization, the simulation degree and cluster control efficiency of the social agent are improved. The behaviors of the intelligent agent are more diversified and natural, and user interaction can be simulated more truly. The cluster control mechanism is rapid and flexible in response, data synchronization and task distribution are more efficient, the problems that in the prior art, intelligent agent behaviors are single, cluster control is slow, data processing is complex, and real-time feedback cannot be achieved are solved, and the information spreading efficiency and the activity degree of the social platform are improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence, distributed systems, and network communications, and in particular to a social intelligent body device and a cluster control method and device thereof. Background Art

[0002] With the rapid development of the Internet and social media, information dissemination has gradually entered the Web3.0 era. The application of big data and intelligent algorithms has greatly promoted the efficiency of information acquisition and dissemination. Social platforms such as Weibo, Twitter, Facebook, Toutiao, etc. have become important channels for users to obtain information, share opinions and participate in interactions. As an application based on artificial intelligence technology, social agents can simulate the behavior of real users on these platforms, autonomously perform social tasks such as browsing, commenting, liking, sharing, etc., and improve the efficiency of information dissemination and platform activity.

[0003] The simulation effect of social agents is one of their core advantages. From the perspective of agents, social agents, as "agents" for information dissemination, need to highly simulate the behavior of real users in order to integrate into the platform ecosystem and interact effectively. By imitating human online behavior, social agents can achieve highly similar interactions with real users and enhance the dissemination effect of information. However, there are some obvious shortcomings in social agents in the prior art. First, the behavior patterns of existing social agents are relatively simple and lack sufficient flexibility and diversity, resulting in low simulation. The behavior of agents is often too mechanical and cannot fully simulate complex user behavior, which limits their scope of application.

[0004] Secondly, the efficiency and flexibility of social agent cluster control are poor. At present, most social agent clusters lack rapid response mechanisms and efficient task scheduling capabilities, and cannot adjust tasks and resource allocation in real time, resulting in low efficiency when performing complex tasks. In addition, data synchronization and communication delay problems still exist, and the real-time feedback capabilities of the agents are insufficient, which affects their performance on large-scale platforms.

[0005] Therefore, improving the fidelity of social agents and the efficiency of cluster control is the main challenge facing current technology. In order to improve the fidelity, it is necessary to make the behavior of social agents more diverse and natural, so as to achieve a higher level of interactive simulation. At the same time, optimizing the task scheduling and data synchronization capabilities of the agent cluster and enhancing its adaptability and flexibility in complex social environments are also the key to improving technical performance. Solving these problems will help the widespread application of social agents in information dissemination. Summary of the invention

[0006] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provides a social intelligent body device and a cluster control method and device thereof.

[0007] In a first aspect, the present invention provides a social agent cluster control method, comprising:

[0008] S1. Count different groups of Internet platform users and build corresponding user profiles;

[0009] S2. Based on the characteristic profiles of Internet users, the user profiles of the agents deployed on the cluster are developed to obtain agents with preferences;

[0010] S3. Creating information consumption activity tasks for the agent based on the BS (Browser / Server, BS for short);

[0011] S4. Collect information consumption data generated by the activities of intelligent agents.

[0012] Optionally, the statistics of different groups of Internet platform users described in step S1 to construct corresponding user feature portraits include: performing multi-dimensional data analysis by obtaining users' historical behavior data, access trajectories, content browsing records, and social interaction information. The analysis dimensions may include, but are not limited to: users' interests, professional information, social circles, purchasing preferences, device usage habits, etc. Then, based on the basic attributes of users such as geographic location, age group, gender, income level, combined with the significant features in their historical behavior, a segmented portrait of the user group is formed. By classifying and aggregating users according to multiple feature dimensions, high-value user groups, potential interest groups, etc. can be identified. Finally, for different user groups, user portrait features for different groups are constructed.

[0013] Optionally, in step S2, based on the Internet user feature profile, the intelligent agent deployed on the cluster is trained to develop a user profile to obtain an intelligent agent with preferences. The specific process is as follows:

[0014] S201. Data collection and preprocessing: Through the user portrait set by the user, extract time series data from the sliding interaction behavior, touch interaction behavior and browsing interaction behavior of the Internet user. At the same time, obtain the user's historical data and user profile data, and use these data as the input source of the user feature portrait to provide comprehensive data support for subsequent steps. After that, the collected behavior sequence data is cleaned and standardized, outliers are removed, and arranged in chronological order to form a complete time series feature data set. For missing values ​​in the data, interpolation or other appropriate methods are used to fill them to ensure the continuity and availability of the time series data.

[0015] S202. Behavior sequence modeling: construct time series feature models for sliding interaction, touch interaction, and browsing interaction behaviors respectively. Specifically, a recursive neural network (RNN) structure is used to model the above behavior sequences to capture the temporal characteristics and long-term dependencies of user behavior data. In order to alleviate the gradient vanishing problem that may occur during RNN network training, a gated recurrent unit (GRU) is further used. The calculation formula of GRU is as follows:

[0016] Update gate: r t =σ(W er e t +W hr h t-1 +b r ) (1)

[0017] Reset gate: z t =σ(W ez e t +W hz h t-1 +b z ) (2)

[0018] Candidate hidden states:

[0019] where e t is the embedding vector of the t-th action, σ is the sigmoid function, and ⊙ is the element-wise product operator.

[0020] In order to support dynamic updates, this application introduces a dynamic weight adjustment mechanism to dynamically update the hidden state through a time decay function and an update increment.

[0021] In order to reduce the impact of outdated behavior on the current user profile, the following time decay formula is designed:

[0022] w t =α.e -β·(T-t) (4)

[0023] Where T is the current timestamp, t is the behavior timestamp, α and β are adjustable parameters used to control the decay speed of the time weight. The time decay weight is used to reduce the impact of historical behavior on the current decision and ensure that the model can give priority to recent behavior.

[0024] Dynamic hidden status update:

[0025]

[0026] where h t is the current hidden state, h t-1 is the previous hidden state. is the candidate hidden state of the current behavior, calculated by the core formula of GRU. t It is the time decay weight, which controls the balance between historical behavior and current behavior.

[0027] The current hidden state is determined by the historical hidden state and the current behavior state, but the contribution of the historical state decays over time. At this point, the model can dynamically adjust the hidden state according to the time weight to balance the influence of short-term and long-term interests.

[0028] S203. Dynamic update mechanism of user portrait: This mechanism includes three aspects: real-time preference prediction, long-term preference trend update, and global user portrait update. The core of this mechanism is to form the dynamic adjustment capability of user portrait by integrating real-time preference prediction and long-term preference trend. Compared with traditional static portrait, this mechanism can adaptively update user portrait and enhance the responsiveness of intelligent agent to changes in user interests.

[0029] Real-time preference prediction:

[0030] P real-time =softmax(W r ·h T +b r ) (6)

[0031] Where P real-time represents the real-time preference vector, represents the user's current interest distribution, W r The weight matrix representing the real-time preference prediction is used to convert the hidden state into preference probability. T is the hidden state of the latest time step, reflecting the user’s current behavior characteristics. r is a bias term used to adjust the baseline value of preference prediction.

[0032] At this point, the intelligent agent predicts the user's preference distribution for various interest tags in real time based on the latest behavioral data.

[0033] Long-term preference trend update:

[0034]

[0035] is the updated long-term preference vector, is the long-term preference vector of the previous time step, P real-time is the current real-time preference vector, and γ is the long-term preference update weight (ranging from 0 to 1), which is used to balance the real-time preference with the historical trend.

[0036] Long-term preference trends are gradually adjusted based on current prediction results to ensure that the portrait can reflect the user's immediate interests while maintaining long-term stability.

[0037] Global user profile update:

[0038] P global =δ·P real-time +(1-δ)·P long-term (8)

[0039] P global It is a global user portrait that integrates short-term interests and long-term trends as the final basis for the agent's response. δ is the fusion weight (ranging from 0 to 1), which adjusts the ratio of short-term preference to long-term preference.

[0040] Through global profiling, the intelligent agent can integrate short-term and long-term interests to provide users with more accurate response services.

[0041] S204. Hierarchical feature fusion: The discussion is divided into two categories, one is intra-sequence feature extraction (using the attention mechanism), and the other is cross-sequence feature fusion (using the cross-view attention mechanism).

[0042] In-sequence feature extraction: Use the in-sequence attention mechanism for each behavior sequence separately to calculate the importance weight of each action within the behavior sequence, thereby identifying the key actions in the user's behavior sequence. Specifically, the attention score calculation formula is:

[0043]

[0044] in is the final output state of the forward GRU model, is the t-th output state of the bidirectional GRU model, The attention score a calculated by this formula t Can reflect action t With current action The relationship between the two-way GRU and the attention score vector is then element-wise multiplied to obtain the output of the attention layer within the sequence.

[0045] Cross-sequence feature fusion: Use the cross-sequence attention mechanism to complete the interaction feature fusion between different behavior sequences. s and browse interactive views b For example, the cross-view attention mechanism This mechanism can effectively capture the correlation and interaction patterns between asynchronous behavior sequences and obtain the output of the cross-view attention layer.

[0046] S205. Multi-task optimization: A multi-task learning framework is designed to improve the recognition of user behavior patterns by simultaneously predicting short-term reading intention and long-term reading intention. The feature embedding vectors of different behavior sequences are fused through splicing operations and then input into the multi-task learning network. The loss functions of the short-term reading intention prediction task and the long-term reading intention prediction task are defined separately.

[0047] Loss function L for short-term reading intention prediction task short :

[0048]

[0049] Loss function L for long-term reading intention prediction task long :

[0050]

[0051] Where y is the actual reading label, N is the total number of samples in the dataset, is the training set.

[0052] The global loss function L global is defined as the weighted sum of the two task losses to enable multi-task learning:

[0053] L global =λL short +(1-λ)L long (12)

[0054] Here, λ is a hyperparameter used to balance the importance of the two tasks.

[0055] In this way, the model can simultaneously learn to predict the robot's immediate reading behavior and long-term reading intention, which not only improves the model's understanding of the robot's behavior, but also enhances its adaptability and accuracy in practical applications. This multi-task learning approach enables the model to capture the complexity and diversity of the robot's behavior, thus playing a key role in information consumption and social interaction.

[0056] S206. Generation and deployment of agent preference features: Combine the real-time preference prediction results with the persistent intent prediction results to form a dynamic update mechanism for user portraits. Through the reverse update mechanism, the generated user portrait features are continuously applied to the agents on the cluster, and finally the agent preference features are cultivated. Then the agent with preferences is deployed to the cluster environment, where the generated agent preference features are used to respond to the user's interactive behavior in real time, and the preference features are dynamically adjusted according to the changes in user behavior data to improve the agent's responsiveness and adaptability.

[0057] Optionally, the creation of information consumption activity tasks for the agent based on the BS architecture described in step S3 includes: creating and managing tasks in a user-friendly Web interface, and using the powerful processing capabilities of the server to assign and execute tasks. The user creates a new agent task in the intuitive task management interface and selects an agent suitable for performing the task from a predefined agent pool. During the task creation process, the user needs to set the task parameters in detail, including but not limited to the target platform, the scheduled execution time, the specific description of the task, and the expected results. In addition, the user can also select a specific type of agent to execute according to the special needs of the task, such as overseas agents customized for different regions or language environments.

[0058] Optionally, the collection of information consumption data generated by the agent activity in step S4. includes: recording various types of information data received by the agent in the process of accessing the Internet media platform, such as the information display in the homepage, recommendation, classification, and search page, including content type, source, release time, and interaction volume (likes, comments, forwarding, etc.). For the content that is browsed in detail, collect comment-related data, including comment content, author, time, interaction volume, etc. The agent controller is responsible for receiving and storing this data, and recording metadata based on the collection behavior, such as collection time, source page, and classification basis, for subsequent analysis and query. After collecting the information consumption data generated by the agent activity, the key to the control method lies in the verification of agent feedback and task completion. The standard for task completion not only includes the integrity of data collection, but also needs to evaluate whether the agent has achieved the predetermined goal according to the task requirements. For example, if the task is to browse a certain type of content, the agent needs to browse effectively according to the task goal (such as a specific content type, time period, interaction volume, etc.), and feedback the results of the task execution, including whether the predetermined interaction volume or browsing depth is reached. By comprehensively analyzing the agent's task execution process and feedback data, the system can automatically determine whether the task is completed, further adjust task parameters or reallocate tasks to ensure the accuracy and efficiency of task execution, thereby achieving all-round control of the agents in the cluster.

[0059] In the second aspect, the present invention provides a social intelligent agent device, including an intelligent agent module, a web automation robot based on Autojs or Jxbrowser, distributed on various public and private cloud hosts or mobile phones. An automatic program that can browse, read, follow, comment, like, and other information consumption and social behaviors in a specified social media based on the assigned task information, and has certain text classification, content monitoring, and image recognition capabilities. A task assignment module that assigns tasks to a specific one or a group of social intelligent agent programs according to execution time and related attributes. A task collection module that collects relevant task feedback and data. An automatic program that handles related errors in task execution. A message intermediary module that includes a task queue unit and a task feedback unit to decouple the relationship between the intelligent agent and the task distributor. Different types of tasks are distributed through a single-point queue or a batch distribution topic, and social intelligent agent feedback is performed through a queue. A data storage module: stores social intelligent agent task information, feedback information, browsing behavior information, social intelligent agent status information, intelligent agent Profile information, etc.

[0060] After the user publishes a task, the task information will be stored in the task information table in the data storage module. The task assignment module reads the task information and feeds the task to the task queue unit in the message mediation module. The agent module obtains the task from the task queue unit and executes it. After the execution is completed, the task result is fed to the feedback queue unit in the message mediation module. The task collection module reads the task feedback information in the feedback queue unit and stores the feedback information, browsing behavior information, profile information, social agent information and other data in the data storage module. In addition, the agent module will separately back up and store all the information in the data storage module.

[0061] In a third aspect, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of the implementation modes of the first and second aspects above is implemented.

[0062] The advantages of the present invention are: through the cultivation of intelligent agent portraits and the optimization of task scheduling, the realism and cluster control efficiency of social intelligent agents are improved; the behavior of intelligent agents is more diverse and natural, and can simulate user interactions more realistically; the cluster control mechanism responds quickly and flexibly, and data synchronization and task allocation are more efficient, which solves the problems of single intelligent agent behavior, slow cluster control, complex data processing, and inability to provide real-time feedback in the prior art, and improves the efficiency of information dissemination and the activity of social platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0064] Figure 1 is a flow chart of the method of the present invention;

[0065] Figure 2 is a structural diagram of a social agent device of the present invention;

[0066] Figure 3 is a logic flow chart of a task distributor in a social agent device of the present invention;

[0067] Figure 4 It is a logical flow chart of the task collector in the social intelligence device of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0069] Example 1

[0070] like Figure 1 , this embodiment relates to a social agent cluster control method, comprising the following steps:

[0071] Step 101: Count different groups of Internet platform users and construct corresponding user feature portraits. In this embodiment, through the refined analysis and division of user group characteristics, the behavioral habits and preferences of different user groups are effectively identified, providing a solid foundation for subsequent data mining and precise push. Specifically, the system first collects multi-dimensional data generated by users on the platform, which includes but is not limited to users' basic attributes, historical operation behaviors and interaction data. Through these data, the user's usage and characteristics on the platform can be fully and systematically grasped.

[0072] As an optional implementation method of this embodiment, first of all, the attribute distribution information of platform users is the core foundation of portrait construction. In order to accurately construct user feature portraits, the system obtains the user's geographic distribution attributes, age group divisions, gender classifications, activity time attributes, and content preference attributes through automated collection and data mining technology. Specifically, geographic distribution attributes can be identified based on the user's IP address or other geographic information to help the platform understand the activity level and content needs of users in different regions; age attributes are divided into different age groups according to the user's registration information or social information, such as minors, youth groups, middle-aged groups, and elderly groups; in terms of gender attributes, the system not only distinguishes between traditional male and female users, but also introduces mixed gender options based on the composition of the platform user group to more accurately reflect the actual user situation.

[0073] As an optional implementation method of this embodiment, in addition to basic demographic characteristics, the activity time attributes of platform users are also key analysis indicators. The system constructs the user's activity time distribution model and activity interval distribution model by analyzing the user's historical behavior data. The activity time distribution model records the user's activity in different time periods, such as morning, afternoon, and evening usage, which can help identify the user's peak usage period; the activity interval distribution model records the interval time between each user's visit to the platform, helping the system identify the user's usage frequency and active cycle.

[0074] As an optional implementation method of this embodiment, the acquisition of content preference attributes provides a richer dimension for the construction of user portraits. In the content preference attributes, the system extracts the information categories that users are interested in by analyzing the content that users browse, click, comment on and share. These categories can be divided according to the classification standards pre-defined by the platform, such as news, entertainment, technology, etc. Each user's content preferences will be sorted according to their level of activity in these different categories to form a content preference portrait of the user. In addition, the system further refines the content classification by extracting keywords that users pay attention to during browsing and interaction. Keywords not only include content-related identification words, but may also involve information such as geographic location and gender preference, making the portrait construction more three-dimensional and comprehensive.

[0075] In this embodiment, for the user's content consumption behavior, it is necessary not only to record the user's click behavior, but also to track the user's stay time in each type of information, interaction frequency and other indicators to more comprehensively reflect the user's actual interest in this type of information. For the information that the user browses in detail, the system will further refine the record, obtain behavioral data such as comments, forwarding, and likes, so as to analyze the user's deep engagement with the content. These data are crucial in the construction of feature portraits, and can help the platform identify high-stickiness user groups and potential core user groups, and provide support for subsequent precision marketing and advertising push.

[0076] Step 102: Based on the Internet user feature profile, the intelligent agent deployed on the cluster is trained to develop a user profile, thereby obtaining an intelligent agent with preference. In this embodiment, for the set user profile, the intelligent agent is trained to develop a user profile, that is, through certain preferred reading behaviors or other specific interactions with the recommendation system, the intelligent agent is trained to develop an interest, thereby obtaining an intelligent agent with preference.

[0077] As an optional implementation of this embodiment, user portrait development is performed on the agent to obtain an agent with preferences, including:

[0078] Data collection and preprocessing: Through the user portrait set by the user, extract time series data from the sliding interaction behavior, touch interaction behavior and browsing interaction behavior of the Internet user. At the same time, obtain the user's historical data and user profile data, and use these data as the input source of the user feature portrait to provide comprehensive data support for subsequent steps. After that, the collected behavior sequence data is cleaned and standardized, outliers are removed, and arranged in chronological order to form a complete time series feature data set. For missing values ​​in the data, interpolation or other appropriate methods are used to fill them to ensure the continuity and availability of the time series data.

[0079] Behavior sequence modeling: Time series feature models are constructed for sliding interaction, touch interaction, and browsing interaction behaviors. Specifically, the recursive neural network (RNN) structure is used to model the above behavior sequences to capture the temporal characteristics and long-term dependencies of user behavior data. In order to alleviate the gradient vanishing problem that may occur during RNN network training, the gated recurrent unit (GRU) is further used. The calculation formula of GRU is as follows:

[0080] Update gate: r t =σ(W er e t +W hr h t-1 +b r ) (1)

[0081] Reset gate: z t=σ(W ez e t +W hz h t-1 +b z ) (2)

[0082] Candidate hidden states:

[0083] where e t is the embedding vector of the t-th action, σ is the sigmoid function, and ⊙ is the element-wise product operator.

[0084] In order to support dynamic updates, this application introduces a dynamic weight adjustment mechanism to dynamically update the hidden state through a time decay function and an update increment.

[0085] In order to reduce the impact of outdated behavior on the current user profile, the following time decay formula is designed:

[0086] w t =α.e -β·(T-t) (4)

[0087] Where T is the current timestamp, t is the behavior timestamp, α and β are adjustable parameters used to control the decay speed of the time weight. The time decay weight is used to reduce the impact of historical behavior on the current decision and ensure that the model can give priority to recent behavior.

[0088] Dynamic hidden status update:

[0089]

[0090] where h t is the current hidden state, h t-1 is the previous hidden state. is the candidate hidden state of the current behavior, calculated by the core formula of GRU. t It is the time decay weight, which controls the balance between historical behavior and current behavior.

[0091] The current hidden state is determined by the historical hidden state and the current behavior state, but the contribution of the historical state decays over time. At this point, the model can dynamically adjust the hidden state according to the time weight to balance the influence of short-term and long-term interests.

[0092] Dynamic update mechanism of user portrait: This mechanism includes three aspects: real-time preference prediction, long-term preference trend update, and global user portrait update. The core of this mechanism is to form the dynamic adjustment capability of user portrait by integrating real-time preference prediction and long-term preference trend. Compared with traditional static portrait, this mechanism can adaptively update user portrait and enhance the responsiveness of intelligent agent to changes in user interests.

[0093] Real-time preference prediction:

[0094] P real-time =softmax(W r ·h T +b r ) (6)

[0095] Where P real-time represents the real-time preference vector, represents the user's current interest distribution, W r The weight matrix representing the real-time preference prediction is used to convert the hidden state into preference probability. T is the hidden state of the latest time step, reflecting the user’s current behavior characteristics. r is a bias term used to adjust the baseline value of preference prediction.

[0096] At this point, the intelligent agent predicts the user's preference distribution for various interest tags in real time based on the latest behavioral data.

[0097] Long-term preference trend update:

[0098]

[0099] is the updated long-term preference vector, is the long-term preference vector of the previous time step, P real-time is the current real-time preference vector, and γ is the long-term preference update weight (ranging from 0 to 1), which is used to balance the real-time preference with the historical trend.

[0100] Long-term preference trends are gradually adjusted based on current prediction results to ensure that the portrait can reflect the user's immediate interests while maintaining long-term stability.

[0101] Global user profile update:

[0102] P global =δ·P real-time +(1-δ)·P long-term (8)

[0103] P global It is a global user portrait that integrates short-term interests and long-term trends as the final basis for the agent's response. δ is the fusion weight (ranging from 0 to 1), which adjusts the ratio of short-term preference to long-term preference.

[0104] Through global profiling, the intelligent agent can integrate short-term and long-term interests to provide users with more accurate response services.

[0105] Hierarchical feature fusion: There are two types of discussion: one is intra-sequence feature extraction (using the attention mechanism), and the other is cross-sequence feature fusion (using the cross-view attention mechanism).

[0106] In-sequence feature extraction: Use the in-sequence attention mechanism for each behavior sequence separately to calculate the importance weight of each action within the behavior sequence, thereby identifying the key actions in the user's behavior sequence. Specifically, the attention score calculation formula is:

[0107]

[0108] in is the final output state of the forward GRU model, is the t-th output state of the bidirectional GRU model, The attention score a calculated by this formula t Can reflect action t With current action The relationship between the two-way GRU and the attention score vector is then element-wise multiplied to obtain the output of the attention layer within the sequence.

[0109] Cross-sequence feature fusion: Use the cross-sequence attention mechanism to complete the interaction feature fusion between different behavior sequences. s and browse interactive views b For example, the cross-view attention mechanism M(v s , v b )=A s (v b , v s , v s )⊙A b (v s , v b , v b in This mechanism can effectively capture the correlation and interaction patterns between asynchronous behavior sequences and obtain the output of the cross-view attention layer.

[0110] Multi-task optimization: A multi-task learning framework is designed to improve the recognition of user behavior patterns by simultaneously predicting short-term and long-term reading intentions. The feature embedding vectors of different behavior sequences are fused through concatenation and then input into the multi-task learning network. The loss functions for the short-term reading intention prediction task and the long-term reading intention prediction task are defined separately.

[0111] Loss function L for short-term reading intention prediction task short :

[0112]

[0113] Loss function L for long-term reading intention prediction task long:

[0114]

[0115] Where y is the actual reading label, N is the total number of samples in the dataset, is the training set.

[0116] The global loss function L global is defined as the weighted sum of the two task losses to enable multi-task learning:

[0117] L global =λL short +(1-λ)L long (12)

[0118] Here, λ is a hyperparameter used to balance the importance of the two tasks.

[0119] In this way, the model can simultaneously learn to predict the robot's immediate reading behavior and long-term reading intention, which not only improves the model's understanding of the robot's behavior, but also enhances its adaptability and accuracy in practical applications. This multi-task learning approach enables the model to capture the complexity and diversity of the robot's behavior, thus playing a key role in information consumption and social interaction.

[0120] Agent preference feature generation and deployment: Combine the real-time preference prediction results with the persistent intent prediction results to form a dynamic update mechanism for user portraits. Through the reverse update mechanism, the generated user portrait features are continuously applied to the agents on the cluster, and finally the agent preference features are cultivated. The agent with preferences is then deployed to the cluster environment, where the generated agent preference features are used to respond to the user's interactive behavior in real time, and the preference features are dynamically adjusted according to changes in user behavior data to improve the agent's responsiveness and adaptability.

[0121] In this embodiment, the user portrait of the intelligent agent can be effectively developed, so that it can complete personalized responses and adjustments based on the user's behavioral characteristics and preferences.

[0122] Step 103: Create tasks for the agent based on the BS architecture.

[0123] In this embodiment, the process of creating tasks for agents based on the BS architecture is managed through an intuitive Web interface, achieving flexible configuration of tasks and efficient scheduling of agents. Users select the target platform, task time, task behavior, and required parameter settings through the Web interface. Users enter the task name and select an agent suitable for performing the task from the predefined agent pool, achieving accurate matching of tasks and agents.

[0124] In this optional implementation, the task management system makes full use of the processing power of the server side, making the allocation and execution of complex tasks efficient and reliable. Users create tasks in the web-based operation interface, select or customize the agent to perform the task, and set the task parameters in detail, including the target platform (such as social media platform, e-commerce platform, etc.), the scheduled task execution time, the specific task description (such as data collection, comment analysis, user behavior tracking, etc.), and the expected task results. For some tasks with special requirements, the system allows users to select specific types of agents according to specific application scenarios, such as agents suitable for different regions and language environments, so as to ensure the effective execution of tasks on a global scale.

[0125] This application further allows users to formulate and assign corresponding tasks based on the characteristics of the tasks, the deployment platform of the agent and the content area of ​​interest, so that the agent can automatically drive the execution. Through this task management mode based on the BS architecture, the task creation, scheduling and monitoring process is more efficient and transparent, and users can flexibly adjust the task execution plan to ensure that the task can be completed accurately, greatly improving the practicality of the system and the execution efficiency of the agent.

[0126] Step 104: Collect information consumption data generated by the agent's activities.

[0127] In this embodiment, the execution process of the task is transparent and controllable by real-time monitoring and data collection of the running status of the agent. During the execution of the task, the system monitors the behavior performance and task completion of the agent in real time, including the progress of the task, the current interaction status, and the performance indicators of the agent. If the system prompts that the task has been completed without error, the user can download all relevant task data with one click.

[0128] To ensure the accuracy and completeness of data collection, the system provides an intuitive task progress bar and status indicator, allowing users to understand the progress of task execution at a glance. Through the real-time monitoring interface, users can clearly view the number of tasks received, the number of tasks completed, the number of errors reported, and other information of the agent. Once the agent completes the task or the task progress bar reaches 100%, the system automatically confirms that the task has been successfully completed and organizes all relevant data generated by the task. These data can be obtained through the one-click download function, and users can quickly obtain task results, log files, analysis reports, etc.

[0129] The collected task data include but are not limited to: project source, project classification label, browsing and collection time, number of likes and comments, keyword identification or classification results of project content, actual project price and payment amount, and other detailed information. After the task is completed, the system stores this data in a structured manner, and users can obtain complete information about the project through the download button, including original HTML code, author ID, number of project collections, number of views, number of shares, project release time, etc. In addition, the system also supports further analysis and statistics of the task data collected by the agent, and users can perform secondary processing on the data according to their needs to obtain deeper analysis conclusions.

[0130] This embodiment provides a safe, efficient and user-friendly intelligent agent management platform through group feature portrait construction, intelligent agent portrait cultivation, real-time task monitoring and one-click download functions, which enhances user experience and system practicality. The above is a method provided by one or more embodiments of this application. Based on the same idea, this application also provides a corresponding social intelligent agent device, reference Figure 2 The schematic diagram contains different module information, among which the logic includes: Agent module, web automation robot based on Autojs or Jxbrowser, distributed on various public and private cloud hosts or mobile phones. An automatic program that can browse, read, follow, comment, like and other information consumption and social behaviors in the specified social media according to the assigned task information, and has certain text classification, content monitoring and image recognition capabilities. Task allocation module, which allocates tasks to a specific one or a group of social agent programs according to execution time and related attributes. Task collection module, which collects relevant task feedback and data. An automatic program that handles related errors in task execution. Message mediation module, which includes task queue unit and task feedback unit, decouples the relationship between agent and task distributor. Different types of tasks are distributed through single-point queues or batch distribution topics, and social agent feedback is performed through a queue. Data storage module: stores social agent task information, feedback information, browsing behavior information, social agent status information, agent Profile information, etc.

[0131] After the user publishes a task, the task information will be stored in the task information table in the data storage module. The task assignment module reads the task information and feeds the task to the task queue unit in the message mediation module. The agent module obtains the task from the task queue unit and executes it. After the execution is completed, the task result is fed to the feedback queue unit in the message mediation module. The task collection module reads the task feedback information in the feedback queue unit and stores the feedback information, browsing behavior information, profile information, social agent information and other data in the data storage module. In addition, the agent module will separately back up and store all the information in the data storage module.

[0132] Example 2

[0133] like Figure 2 ,This embodiment relates to a social agent cluster control device, including: an agent module, a web automation robot based on Autojs or Jxbrowser or, distributed on various public and private cloud hosts or mobile phones; an automatic program that can browse, read, follow, comment, like and other information consumption and social behaviors in designated social media independently according to the assigned task information, and has the capabilities of text classification, content monitoring and image recognition;

[0134] A task allocation module that allocates tasks to a specific social agent program or a group of social agents according to execution time and related attributes;

[0135] Task collection module, which collects relevant task feedback and data; automatic program to handle relevant errors in task execution;

[0136] The message mediation module includes a task queue unit and a task feedback unit, which decouples the relationship between the agent and the task distributor; it distributes different types of tasks through a single-point queue or a batch distribution topic, and provides social agent feedback through a queue;

[0137] Data storage module, storing social agent task information, feedback information, browsing behavior information, social agent status information, and agent Profile information;

[0138] After the user publishes a task, the task information will be stored in the task information table in the data storage module; the task assignment module reads the task information and feeds the task to the task queue unit in the message mediation module; the agent module obtains the task from the task queue unit and executes it, and after execution, the task result is fed to the feedback queue unit in the message mediation module; the task collection module reads the task feedback information in the feedback queue unit, and stores the feedback information, browsing behavior information, Profile information, social agent information and other data in the data storage module; in addition, the agent module will separately back up and store all the information in the data storage module.

[0139] As an optional implementation of this embodiment, the social intelligence module can have 11 behaviors, including recommended browsing, artificial intelligence recommended browsing, homepage browsing, artificial intelligence homepage browsing, category browsing, search browsing, attention, comment, like, reply, and content publishing. Recommended browsing: browsing with preference on the recommended page of the social platform, with preference for artificially given keywords; artificial intelligence recommended browsing refers to browsing with preference on the recommended page of the social platform, with preference for artificially given subject categories, and using the results of classification of the classifier model for judgment and comparison. For example, the artificial intelligence body preference is set to positive energy news, and the positive energy detection classifier is used to judge whether the recommended page news is positive energy news, and then decide whether to click to read. Homepage browsing: browsing with preference on the user homepage of the social platform, with preference for artificially given keywords; artificial intelligence homepage browsing refers to browsing with preference on the user homepage of the social platform, with preference for artificially given subject categories, and using the results of classification of the classifier model for judgment and comparison. For example, the artificial intelligence body preference is set to positive energy news, and the positive energy detection classifier is used to judge whether the recommended page news is positive energy news, and then decide whether to click to read. Search browsing: random browsing on the keyword search page of the social platform, with keywords given manually; category browsing: random browsing on the category topic page of the social platform, with category topics given manually; follow: follow users on the social platform; comment: comment on articles or videos on the social platform; like: like articles or videos on the social platform; reply: reply to a comment under an article or video on the social platform; publish content: publish specified content on the social platform, which can be graphic content or a video generated by artificial intelligence or specified manually.

[0140] As an optional implementation of this embodiment, refer to Figure 3 The task assignment module is responsible for querying the task table and publishing the expired tasks to the message intermediary module in the form of broadcast. After the task is published, the task assignment module enters a short dormant state. At the same time, the task assignment module records the status of the task. The completion of the task needs to meet the following conditions: the number of task profiles received = the number of profiles completed + the number of profiles failed.

[0141] As an optional implementation of this embodiment, refer to Figure 4, the task collection module accepts the social agent's status feedback or task feedback. When the feedback is status feedback, the task collection module will update the social agent's status information and store the updated status information in the data storage module. When the task feedback contains task failure information and the task still has remaining times, the task collection module will change the task status to "pending retry" and require the social agent module to retry. If the task has no remaining times, the task status is changed to "task failed"; if the task is successful and this task is a browsing behavior, the relevant browsing data (such as task browsing time, reading content, etc.) will be recorded and stored in the data storage module; if the task is a non-browsing behavior, the data related to this behavior will be recorded and stored. If the task has a number of cycles, it will enter a loop and repeat the task; if the task has no remaining cycles, the task status will be changed to "task end".

[0142] As an optional implementation method of this embodiment, the message intermediary module includes a task queue unit and a task feedback unit, which is responsible for decoupling the relationship between the social agent and the task distributor, distributing different types of tasks through a single-point queue or batch distribution topics, and providing social agent feedback through a feedback queue; specifically, the message intermediary module acts as a bridge intermediary between the social agent and the task distributor and task collector, and is responsible for indirectly storing relevant data in the data storage module.

[0143] As an optional implementation method of this embodiment, the data storage module serves as the data center of the entire social agent cluster, and mainly stores and protects social agent task information, feedback information, browsing behavior information, social agent status information, and agent Profile information.

[0144] Example 3

[0145] The present application also provides a computer readable medium, which stores a computer program, which can be used to execute the above Figure 1 and the methods and Figure 2 Device provided.

[0146] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute the methods provided by the above-mentioned method embodiments.

[0147] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the above program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment. The above storage medium includes various storage media that can store program codes, such as ROM, RAM, magnetic disk or optical disk.

[0148] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0149] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on this understanding, the above technical solution, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0150] Although the present invention has been described in detail above by general explanation and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.

Claims

1. A social agent cluster control method, characterized in that: include: S1. Count different groups of Internet platform users and build corresponding user profiles; S2. Based on the characteristic profiles of Internet users, the user profiles of the agents deployed on the cluster are developed to obtain agents with preferences; S3. Create information consumption activity tasks for the agent based on the BS architecture; S4. Collect information consumption data generated by the activities of intelligent agents.

2. The social agent cluster control method according to claim 1, characterized in that: According to step S2, based on the Internet user feature profile, the user profile of the intelligent agent deployed on the cluster is cultivated to obtain an intelligent agent with preferences, including: S201 Data collection and preprocessing: Through the user portrait set by the user, extract time series data from the sliding interaction behavior, touch interaction behavior and browsing interaction behavior of the Internet user; at the same time, obtain the user's historical data and user profile data, and use these data as the input source of the user feature portrait to provide comprehensive data support for subsequent steps; then clean and standardize the collected behavior sequence data, remove outliers, and arrange them in chronological order to form a complete time series feature data set; for missing values ​​in the data, use interpolation or other appropriate methods to fill them to ensure the continuity and availability of the time series data; S202. Behavior sequence modeling: construct time series feature models for sliding interaction, touch interaction, and browsing interaction behaviors respectively; specifically, use a recurrent neural network (RNN) structure to model the above behavior sequences to capture the temporal characteristics and long-term dependencies of user behavior data; in order to alleviate the gradient vanishing problem that may occur during RNN network training, a gated recurrent unit (GRU) is further used; the calculation formula of GRU is as follows: Update gate: r t =σ(W er e t +W hr h t-1 +b r ) (1) Reset gate: z t =σ(W ez e t +W hz h t-1 +b z ) (2) Candidate hidden states: where e t is the embedding vector of the t-th action, σ is the sigmoid function, and ⊙ is the element-wise product operator; In order to support dynamic updates, this application introduces a dynamic weight adjustment mechanism to dynamically update the hidden state through a time decay function and an update increment; In order to reduce the impact of outdated behavior on the current user profile, the following time decay formula is designed: w t =a·e -β·(T-t) (4) Where T is the current timestamp, t is the behavior timestamp, α and β are adjustable parameters used to control the decay speed of the time weight. The time decay weight is used to reduce the impact of historical behavior on the current decision, ensuring that the model can give priority to recent behavior; Dynamic hidden status update: where h t is the current hidden state, h t-1 is the previous hidden state; is the candidate hidden state of the current behavior, calculated by the core formula of GRU; w t The time decay weight controls the balance between historical behavior and current behavior; The current hidden state is determined by the historical hidden state and the current behavior state, but the contribution of the historical state decays over time. At this point, the model can dynamically adjust the hidden state according to the time weight to balance the influence of short-term and long-term interests. S203. Dynamic update mechanism of user portrait: This mechanism includes three aspects: real-time preference prediction, long-term preference trend update, and global user portrait update. The core of this mechanism is to form the dynamic adjustment capability of user portrait by integrating real-time preference prediction and long-term preference trend. Compared with traditional static portrait, this mechanism can adaptively update user portrait and enhance the responsiveness of intelligent agent to changes in user interests. Real-time preference prediction: P real-time =softmax(W r ·h T +b r ) (6) Where P real-time represents the real-time preference vector, represents the user's current interest distribution, W r The weight matrix representing the real-time preference prediction is used to convert the hidden state into the preference probability; h T is the hidden state of the latest time step, reflecting the user's current behavior characteristics; b r is a bias term used to adjust the baseline value of preference prediction; At this point, the agent predicts the user's preference distribution for various interest tags in real time based on the latest behavior data; Long-term preference trend update: is the updated long-term preference vector, is the long-term preference vector of the previous time step, P real-time is the current real-time preference vector, γ is the long-term preference update weight (ranging from 0 to 1), which is used to balance the real-time preference with the historical trend; Long-term preference trends are gradually adjusted based on current prediction results to ensure that the portrait reflects the user's immediate interests while maintaining long-term stability; Global user profile update: P global =δ·P real-time +(1-δ)·P long-term (8) P global It is a global user portrait that integrates short-term interests and long-term trends as the final basis for the agent's response; δ is the fusion weight (ranging from 0 to 1), which adjusts the ratio of short-term preferences to long-term preferences; Through global profiling, the agent can integrate short-term and long-term interests to provide users with more accurate response services; S204. Hierarchical feature fusion: There are two types of discussion: one is intra-sequence feature extraction (using the attention mechanism), and the other is cross-sequence feature fusion (using the cross-view attention mechanism); In-sequence feature extraction: The in-sequence attention mechanism is used for each behavior sequence separately to calculate the importance weight of each action in the behavior sequence, thereby identifying the key actions in the user behavior sequence; specifically, the attention score calculation formula is: in is the final output state of the forward GRU model, is the t-th output state of the bidirectional GRU model, The attention score a calculated by this formula t Can reflect action t With current action The relationship between , thereby determining the important actions; then the output of the bidirectional GRU is element-wise multiplied with the corresponding attention score vector to obtain the output of the attention layer within the sequence; Cross-sequence feature fusion: Use the cross-sequence attention mechanism to complete the interaction feature fusion between different behavior sequences; s and browse interactive views b For example, the cross-view attention mechanism IA(v s ,v b )=A s (v b ,v s ,v s )⊙A b (v s ,v b ,v b in This mechanism can effectively capture the correlation and interaction patterns between asynchronous behavior sequences and obtain the output of the cross-view attention layer; S205. Multi-task optimization: A multi-task learning framework is designed to improve the recognition of user behavior patterns by simultaneously predicting short-term reading intention and long-term reading intention; the feature embedding vectors of different behavior sequences are fused through splicing operations and then input into the multi-task learning network; the loss functions of the short-term reading intention prediction task and the long-term reading intention prediction task are defined separately; Loss function L for short-term reading intention prediction task short : Loss function L for long-term reading intention prediction task long : Where y is the actual reading label, N is the total number of samples in the dataset, is the training set; The global loss function L global is defined as the weighted sum of the two task losses to enable multi-task learning: THE global =λL short +(1-λ)L long (12) Here, λ is a hyperparameter used to balance the importance of the two tasks; S206. Generation and deployment of agent preference features: Combine the real-time preference prediction results with the persistent intention prediction results to form a dynamic update mechanism for user portraits; through the reverse update mechanism, continuously apply the generated user portrait features to the agents on the cluster, and finally complete the development of agent preference features; then deploy the agent with preferences to the cluster environment, where the generated agent preference features are used to respond to the user's interactive behavior in real time, and dynamically adjust the preference features according to changes in user behavior data to improve the agent's responsiveness and adaptability.

3. The social agent cluster control method according to claim 1, characterized in that: The step S1 of counting different groups of Internet platform users and constructing corresponding user feature portraits includes: After the creation of the intelligent agent, the user portrait of the intelligent agent is developed according to the set user portrait, and the intelligent agent with preferences is obtained, including: multi-dimensional data analysis is performed by obtaining the user's historical behavior data, access trajectory, content browsing history and social interaction information; the analysis dimensions may include but are not limited to: the user's interests and hobbies, professional information, social circle, purchase preferences, device usage habits, etc.; then, based on the user's basic attributes such as geographic location, age group, gender, income level, combined with the significant features in their historical behavior, a segmented portrait of the user group is formed; by classifying and aggregating users according to multiple feature dimensions, high-value user groups, potential interest groups, etc. can be identified; finally, for different user groups, user portrait features for different groups are constructed.

4. The social agent cluster control method according to claim 1, characterized in that: The creation of information consumption activity tasks for the agent based on the BS architecture described in step S3 includes: creating and managing tasks in a user-friendly Web interface, and using the powerful processing capabilities of the server to assign and execute tasks; the user creates a new agent task in an intuitive task management interface, and selects an agent suitable for performing the task from a predefined agent pool; during the task creation process, the user sets task parameters, including the target platform, the scheduled execution time, the specific description of the task, and the expected results; the user selects a specific type of agent to execute according to the special requirements of the task.

5. The social agent cluster control method according to claim 1, characterized in that ,,The collection of information consumption data generated by the intelligent agent's activities as described in step S4 includes: recording the information data received by the intelligent agent in the process of accessing the Internet media platform, including content type, source, release time, and interaction volume; for the content that is browsed in detail, collecting comment-related data, including comment content, author, time, and interaction volume; the intelligent agent controller receives and stores this data, and records metadata based on the collection behavior for subsequent analysis and query.

6. A social agent cluster control device, characterized in that: include: Intelligent agent module, based on Autojs or Jxbrowser or web automation robot, distributed on various public and private cloud hosts or mobile phones; it is an automatic program that can browse, read, follow, comment, like and other information consumption and social behaviors in designated social media according to the assigned task information, and has the ability of text classification, content monitoring and image recognition; A task allocation module that allocates tasks to a specific social agent program or a group of social agents according to execution time and related attributes; Task collection module, which collects relevant task feedback and data; automatic program to handle relevant errors in task execution; The message mediation module includes a task queue unit and a task feedback unit, which decouples the relationship between the agent and the task distributor; it distributes different types of tasks through a single-point queue or a batch distribution topic, and provides social agent feedback through a queue; Data storage module, storing social agent task information, feedback information, browsing behavior information, social agent status information, and agent Profile information; After the user publishes a task, the task information will be stored in the task information table in the data storage module; The task assignment module reads the task information and feeds the task to the task queue unit in the message intermediary module; the agent module obtains the task from the task queue unit and executes it, and after execution, feeds the task result to the feedback queue unit in the message intermediary module; the task collection module reads the task feedback information in the feedback queue unit, and stores the feedback information, browsing behavior information, Profile information, social agent information and other data in the data storage module; in addition, the agent module will separately back up and store all the information in the data storage module.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the social agent cluster control method described in any one of claims 1 to 4 is implemented.

Images