Interaction method, communication device and computer program product

Abstract

The application discloses an interaction method, a communication device and a computer program product, and belongs to the technical field of intelligent agents. The interaction method comprises the following steps: receiving multi-source data, wherein the multi-source data comprises memory data of an intelligent agent and user feature data of an application platform; performing feature fusion processing on the multi-source data to obtain a user global feature vector, wherein the dimension of the user global feature vector represents a user portrait label type; generating a service instruction based on the dimension in the user global feature vector and an interaction scene, wherein the service instruction is used for instructing the intelligent agent to perform a task corresponding to the interaction scene.

Description

Technical Field

[0001] This application belongs to the field of intelligent agent interaction technology, specifically relating to an interaction method, communication device and computer program product. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent interaction systems have been widely used in scenarios such as intelligent customer service and intelligent agent tasks that interact with large language models (LLM), and users' demand for personalized and adaptive services is becoming increasingly prominent.

[0003] In related technologies, in the field of intelligent agents and intelligent interaction, the terminal usually collects intelligent agent data and then uses the intelligent agent data to serve users. For example, intelligent customer service systems rely solely on dialogue records to judge user needs, which leads to bias in demand identification and fails to meet users' increasingly sophisticated interaction needs. Summary of the Invention

[0004] This application aims to provide an interaction method, communication device, and computer program product that at least solves the problem in related technologies where relying solely on intelligent agent data to serve users leads to biases in demand identification.

[0005] In a first aspect, embodiments of this application propose an interaction method, including: Receive multi-source data, including the agent's memory data and the application platform's user characteristic data; The multi-source data is subjected to feature fusion processing to obtain a global user feature vector, and the dimension of the global user feature vector represents the user profile label type. Based on the dimensions in the user's global feature vector and the interaction scenario, a service instruction is generated. The service instruction is used to instruct the agent to perform a task corresponding to the interaction scenario.

[0006] Secondly, embodiments of this application provide an interaction method, including: Acquire the memory data of the intelligent agent and the user characteristic data of the application platform; Send the memory data and the user feature data to the cloud; The agent receives service instructions from the cloud, which instruct the agent to perform tasks corresponding to the interaction scenario.

[0007] Thirdly, embodiments of this application provide a communication device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect, or the steps of the method described in the second aspect.

[0008] Fourthly, embodiments of this application provide a computer program product stored in a storage medium, which is executed by at least one processor to implement the steps of the method described in the first aspect, or the steps of the method described in the second aspect.

[0009] In the embodiments of this application, multi-source data is first received, including the agent's memory data and the application platform's user feature data. Then, this multi-source data undergoes feature fusion processing to obtain a global user feature vector. The dimensions of the global feature vector represent the user profile label type. Finally, based on the dimensions in the global user feature vector and the interaction scenario, a service instruction is generated. This service instruction instructs the agent to perform a task corresponding to the interaction scenario. This embodiment of the application can fuse received multi-source data. Compared to related technologies that only use agent data to serve users, the data is more diverse. The fused global user feature vector results in diverse profile dimensions, and combined with the scenario, it can provide users with more personalized needs, improving service accuracy. Attached Figure Description

[0010] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of an interaction method provided in one embodiment of this application; Figure 2 This is a service decision diagram provided in one embodiment of this application; Figure 3 This is a flowchart of another interaction method provided in one embodiment of this application; Figure 4 This is a schematic diagram of a user interface provided in one embodiment of this application; Figure 5 This is a schematic diagram of the interaction process provided in one embodiment of this application; Figure 6 This is a schematic diagram of an interactive device provided in one embodiment of this application; Figure 7 This is a schematic diagram of another interactive device provided in one embodiment of this application; Figure 8 This is a schematic diagram of a communication device provided in one embodiment of this application. Detailed Implementation

[0011] Embodiments of the present invention will now be described in detail. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0012] The terms "first" and "second" in the specification and claims of this application may explicitly or implicitly include one or more of the features. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0014] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0015] In related technologies, agent memory and user feature data on the application platform are isolated from each other. Agent memory is stored only locally on a single agent, lacking a standardized channel to transmit valuable information to user feature data. User feature data relies solely on in-application behavior updates, lacking a linkage and optimization mechanism with agent interaction data. The two types of data lack clear interaction paths and dynamic collaborative optimization rules, resulting in a single-dimensional profile and wasted memory value. Ultimately, this makes it difficult to improve the decision-making accuracy and personalization of the edge-side intelligent interaction system, failing to meet users' increasingly sophisticated interaction needs.

[0016] Therefore, the interaction method, communication device, and computer program product provided in this application can more accurately identify users' potential needs by simultaneously collecting intelligent agent interaction memory data and user feature data of the application platform, and fusing the data to construct a comprehensive and dynamic user profile.

[0017] The following is combined Figures 1-8 The present application provides a detailed description of an interactive method, communication device, and computer program product through specific embodiments and application scenarios.

[0018] like Figure 1 The diagram shown is a flowchart of an interaction method provided in an embodiment of this application. Figure 1 As shown, the interaction method may include the content shown in steps 101 to 103.

[0019] In S101, multi-source data is received, including the agent's memory data and the application platform's user feature data.

[0020] The memory data of the intelligent agent may include data such as user operation feedback, task execution trajectory, and interactive content annotation collected in real time during the execution of tasks by establishing a real-time connection with various functional intelligent agents through the application programming interface (API). It may also include data generated during the execution of tasks by other intelligent agents, subject to actual application, and this embodiment does not limit it.

[0021] User characteristic data may include user operation behavior, frequency of function use, device usage parameters, and other data that can reflect user characteristics, depending on the actual application. This embodiment does not limit it.

[0022] In S102, feature fusion processing is performed on multi-source data to obtain a global user feature vector. The dimension of the global user feature vector represents the user profile label type.

[0023] The user global feature vector can include multiple dimensions. Feature categories can be divided according to the dimensional distribution of the feature vector. Each feature category corresponds to a user profile label type. For example, by combining the interaction records of the agent's memory nodes with the usage preferences of the feature attribute nodes, interaction-function preference category labels can be generated.

[0024] The dimension of the vector is related to the received data. This application does not limit the number of dimensions, but the actual application shall prevail.

[0025] In S103, service instructions are generated based on the dimensions in the user's global feature vector and the interaction scenario. These service instructions are used to instruct the agent to perform tasks corresponding to the interaction scenario.

[0026] The interaction scenarios may include intelligent customer service scenarios, intelligent education service scenarios, intelligent office collaboration scenarios, etc., and may also include other interaction scenarios with intelligent agents. This embodiment does not limit these scenarios and the actual application shall prevail.

[0027] In the embodiments of this application, multi-source data is first received, including the agent's memory data and the application platform's user feature data. Then, this multi-source data undergoes feature fusion processing to obtain a global user feature vector. The dimensions of the global feature vector represent the user profile label type. Finally, based on the dimensions in the global user feature vector and the interaction scenario, a service instruction is generated. This service instruction instructs the agent to perform a task corresponding to the interaction scenario. This embodiment of the application can fuse received multi-source data. Compared to related technologies that only use agent data to serve users, the data is more diverse. The fused global user feature vector results in diverse profile dimensions, and combined with the scenario, it can provide users with more personalized needs, improving service accuracy.

[0028] In one possible implementation of this application, performing feature fusion processing on multi-source data to obtain a user global feature vector may include: constructing a knowledge graph based on multi-source data; and performing feature fusion on the knowledge graph based on a graph neural network model to obtain a user global feature vector.

[0029] The knowledge graph construction process involves filling the nodes of an initial knowledge graph with multi-source data to obtain the knowledge graph. That is, an initial knowledge graph can be pre-set, which may contain preset nodes, such as user nodes, data source nodes, feature attribute nodes, interaction scenario nodes, and agent memory nodes. The type of nodes can be determined based on historical data or actual conditions, and the preset node types are subject to actual application; this embodiment does not impose any limitations. Simultaneously, basic attribute fields can be set for each node. For example, an agent memory node includes attribute fields such as a unique memory document (ID), memory type, memory content summary, memory generation timestamp, and associated agent identifier. Other nodes are configured with attributes according to functional requirements. For example, a user node may include attribute types such as a unique identifier and feature attribute nodes. Attribute fields may also include others; this embodiment will not list them all, and the actual application shall prevail.

[0030] The architecture of the graph neural network (GNN) model can be determined based on the knowledge graph. It can be a graph convolution network (GCN), a graph attention network (GAT), or other graph neural networks, depending on the actual application. This embodiment does not limit the choice.

[0031] In this embodiment, a knowledge graph is constructed by multi-source data, which can map the correspondence between different data to graph nodes, making it easier to determine the relationship between different data. Then, based on the GNN model, the features of each data are fused to obtain a multi-dimensional global feature vector of the user. Each dimension corresponds to a user profile label type, which can then determine the current user profile and serve the user according to their needs.

[0032] In one possible implementation of this application, before constructing a knowledge graph based on multi-source data, the interaction method may further include: preprocessing memory data and user feature data to obtain standardized data, wherein the preprocessing includes at least one of the following: data cleaning, format conversion, and data reduction.

[0033] In this embodiment, after receiving multi-source data, the cloud engine layer can preprocess the multi-source data. This involves data cleaning to remove blank, duplicate, and abnormal data, such as values ​​exceeding reasonable ranges and logically contradictory records. Then, data conversion is used to unify data of different formats and units into a standard format recognizable by the cloud, such as unified timestamp specifications and data units. Finally, data reduction is used to reduce the amount of data while retaining key information. Ultimately, this provides high-quality, standardized input data for subsequent data fusion, solving the problems of chaotic data formats and low validity in related technologies, and providing high-quality, standardized input for subsequent data fusion.

[0034] In one possible implementation of this application, constructing a knowledge graph based on multi-source data may include: determining nodes in the knowledge graph based on memory data and user feature data, wherein each node includes preset attributes; and filling each node in the knowledge graph with agent memory data and user feature data according to preset attributes and correspondences to obtain the knowledge graph.

[0035] In this embodiment, based on the preprocessed data from the above embodiments, the node types in the knowledge graph are determined, which may include user nodes, data source nodes, feature attribute nodes, interaction scenario nodes, and agent memory nodes. Basic attribute fields can also be preset for each node. For example, agent memory nodes may include basic attribute fields such as memory ID, memory type, memory content summary, memory generation timestamp, and associated agent identifier. Other nodes can be configured with attributes according to functional requirements. For example, user nodes may include attribute fields such as unique identifier and feature attribute node. Then, the two types of data are mapped to the graph nodes according to their corresponding relationships.

[0036] For example, agent memory data is populated into the attribute fields of agent memory nodes. Agent memory nodes are connected to user nodes through memory ownership edges, to interaction scene nodes through scene matching edges, and to data source nodes through source association edges.

[0037] User feature data is populated into the attribute fields of feature attribute nodes. Feature attribute nodes are connected to user nodes through feature edges and to data source nodes through data tracing edges. All nodes work together to form the initialization of the knowledge graph.

[0038] In this embodiment, a knowledge graph is constructed using multi-source data, which can map the correspondence between different data to graph nodes, making it easier to determine the association between different data.

[0039] In one possible implementation of this application, feature fusion of a knowledge graph based on a graph neural network model to obtain a user global feature vector may include: determining a graph neural network model based on the number of nodes and edge density of the knowledge graph; converting the data in the knowledge graph into a feature matrix, the feature matrix including node features; using the graph neural network model to aggregate and propagate the feature matrix to obtain a target knowledge graph, the target knowledge graph including the node attributes and edge weights of each node; and using the graph neural network model to perform feature fusion of the target knowledge graph to obtain a user global feature vector.

[0040] After the knowledge graph is initialized, a suitable GNN model structure can be selected based on the number of nodes and edge connection density of the knowledge graph, and the model hierarchy and hidden layer dimensions can be determined.

[0041] In one example, the core parameters of the model can be initialized. These core parameters may include node embedding dimension, initial edge weights, activation function type, and iteration learning rate. The training objectives are to optimize the accuracy of inter-node association weight calculation and improve the accuracy of global association inference. The initial graph data is then transformed into a model input format. The graph data may include all nodes. The transformed model input format is a feature matrix, for example, a feature matrix containing agent memory nodes, an index matrix containing all edge connections, such as an index matrix containing agent memory node associated edges, and an edge attribute matrix, ensuring that the data format meets the model requirements.

[0042] After initializing the core parameters of the model, the GNN model is started. Through multi-layer convolution or attention mechanism, the features of all nodes in the graph are aggregated and propagated to obtain the target knowledge graph including the node attributes and edge weights of each node. Then, the graph neural network model is used to fuse the features of the target knowledge graph to obtain the user global feature vector containing the core information of the two types of data.

[0043] In this embodiment, feature fusion of various data is completed based on the GNN model to obtain a multi-dimensional global feature vector of users. Each dimension corresponds to a user profile label type, which can then determine the current user profile and serve users according to their needs.

[0044] In one possible implementation of this application, a graph neural network model is used to aggregate and propagate the feature matrix to obtain a target knowledge graph. This may include: inputting node features into the graph neural network model for aggregation and propagation to obtain bidirectional features between multiple nodes; generating low-dimensional dense embedding vectors for each node based on the feature aggregation results using the graph neural network model; calculating the association strength between any two nodes based on a preset algorithm; and obtaining the target knowledge graph based on the bidirectional features between multiple nodes, the low-dimensional dense embedding vectors of each node, and the association strength between any two nodes.

[0045] In this embodiment, after starting the GNN model, the features of all nodes in the graph are aggregated and propagated through multi-layer convolution or attention mechanisms. The attribute features of the agent's memory nodes, such as memory type and memory content summary, are combined with the user features of the feature attribute nodes, such as function usage frequency and behavioral preferences, as input. These features are then passed to user nodes, interaction scene nodes, and other nodes through their respective associated edges, forming a bidirectional feature interaction between multiple nodes. Based on the feature aggregation results, the model generates low-dimensional dense embedding vectors for all nodes, ensuring that the vectors of different types of nodes are in the same space.

[0046] Then, the association strength between any two nodes is calculated using algorithms such as cosine similarity and Euclidean distance, which is the weight of the key edge between the two nodes. For example, the weight of key edges such as agent memory node-user node, agent memory node-feature attribute node, and feature attribute node-interaction scene node are calculated. The calculation results are updated to the corresponding edge attributes in real time to form a dynamic association weight matrix.

[0047] Finally, based on the bidirectional features between multiple nodes determined above, the low-dimensional dense embedding vectors of each node, and the association strength between any two nodes, the target knowledge graph is obtained.

[0048] This embodiment establishes a mapping relationship between agent memory data and user feature data, which can mine information such as user needs and behavioral patterns reflected in agent memory. This information is then deeply integrated with user basic characteristics and historical behavior data collected from the application platform to construct a more comprehensive and three-dimensional user profile. This solves the problem of data isolation and difficulty in fully depicting users in traditional systems, and provides a data foundation for precise services.

[0049] After obtaining the target knowledge graph in the above embodiments, there may be problems such as low correlation between nodes and logical contradictions in the attributes of nodes. In this case, it is necessary to adjust the problematic nodes or data, such as by tracing historical data and analyzing the causes, to adjust the relationship between each node in order to ensure that the overall correlation logic of the knowledge graph is reasonable.

[0050] In one possible implementation of this application, the interaction method may further include: determining that there is a conflict between nodes when nodes meet a first preset condition, wherein the first preset condition includes at least one of the following: the association strength between any two nodes is less than a first preset threshold, the same node has different node attributes, and the data between different nodes contradict each other; in the case of a conflict between nodes, analyzing the cause of the conflict based on the association path of the conflicting nodes to obtain an attribution result; and adjusting the attributes of the conflicting nodes based on the attribution result to obtain updated node attributes and edge weights.

[0051] In this embodiment, association conflict determination rules can be set based on the association strength matrix output by the GNN and the global inference results. For example, when the association strength between any two nodes is lower than a preset lower limit, or when there is a logical contradiction in the node attributes, it is determined to be a conflict. For example, the memory generation scenario of the agent memory node is inconsistent with the attributes of the interaction scenario node, or the behavioral data of the feature attribute node contradicts the interaction record of the agent memory node. The global inference capability of the GNN can be invoked to trace back the association path of the conflicting nodes, analyze the cause of the conflict, such as the difference in data collection time, source deviation, etc., and output an attribution report. Based on the attribution results, a reconciliation operation is performed. If the conflict is due to scenario difference, the association weight between the corresponding node and the interaction scenario node is adjusted. If the conflict is due to data source deviation, the attribute value of the conflicting node is adjusted in combination with the credibility of the data source node. Then, the graph node attributes and edge weights are updated according to the reconciled data or nodes to ensure that the overall association logic is reasonable.

[0052] In one example, based on the harmonized graph, GNN node feature fusion computation is initiated, aggregating the embedding vectors of agent memory nodes, feature attribute nodes, user nodes, etc., to generate a global user feature vector containing two types of core data information. Feature categories are divided according to the dimensional distribution of the feature vectors, with each category corresponding to a user profile tag type. Dimensions with a weight ratio higher than a preset threshold are extracted from the vectors and transformed into tag content. For example, combining the interaction records of agent memory nodes and the usage preferences of feature attribute nodes, interaction-functional preference tags are generated. Association attributes are added to each tag, including the association path of the corresponding node, GNN inference confidence, and data source identifier. The tags are connected to user nodes through tag edges to complete the binding of tags to the graph.

[0053] If new data is received from the terminal, this embodiment can use incremental training and adaptive decision-making to update the feature dimensions using the new data, without the need for full reconstruction.

[0054] In one possible implementation of this application, the interaction method may further include: updating the knowledge graph based on the received data when the received data meets a second preset condition, wherein the second preset condition includes at least one of the following: new memory data uploaded by the terminal, new user feature data, and user feedback data.

[0055] In this embodiment, when new memory data, new user feature data, user feedback data, and other new data are received from the terminal, an incremental update mechanism can be adopted to load the basic parameters of the previous version and update only the feature dimensions involved in the fusion label. There is no need for full reconstruction. Compared with related technologies, this can significantly reduce the computational power consumption of model training and significantly improve update efficiency.

[0056] In one example, the trigger conditions for knowledge graph updates can be set: when new agent memory data, new user feature data, or user feedback data are received, the update process is triggered. When new data is received, each data point is updated to the corresponding node. For example, agent memory data is updated to the agent memory node to add nodes or supplement attributes, user feature data is updated to the feature attribute node, and the node feature aggregation is repeated to the label generation stage. Incremental updates are performed on graph nodes, edge weights, feature vectors, and labels. Based on the feedback data, the GNN inference parameters are adjusted. If the feedback shows label deviation, the corresponding node is optimized, such as optimizing the feature aggregation weights or attributes of the agent memory node, and the label content is updated synchronously to ensure that the graph and labels always adapt to changes in user needs.

[0057] In one example, cross-validation and backtracking optimization logic can be introduced to monitor the model fit in real time, automatically adjust the regularization parameters to avoid overfitting, ensure that the model accuracy is maintained at a high level, and solve the problem of poor generalization ability of related technical models.

[0058] In one possible implementation of this application, generating service instructions based on the dimensions in the user's global feature vector and the interaction scenario may include: obtaining user profile tags corresponding to dimensions in the user's global feature vector whose weight percentage is higher than a second preset threshold; and generating service instructions based on the user profile tags corresponding to the dimensions in the user's global feature vector whose weight percentage is higher than the second preset threshold and the interaction scenario.

[0059] After obtaining the user's global feature vector, core user features are extracted, such as high-weight demand preference tags and high-frequency behavior patterns. These are then combined with a pre-defined service rule base, such as feature-service mapping relationships and service execution constraints, to perform service matching. Simultaneously, service content details are adjusted based on the user's current interaction scenario, such as interaction time and terminal device type, such as adapting the service display format to the terminal device. After service matching is completed, structured service instructions are generated.

[0060] In one instance, after the service instruction is generated, it can be format-validated to ensure that the instruction structure is complete and the parameters are correct. Then, it is transmitted to the instruction distribution queue to wait for the network transport layer to send it, thus completing the entire process of generating a personalized service instruction.

[0061] This embodiment uses user profile tags with high weighting to determine user needs and adjusts strategies based on the current interaction scenario to generate highly adaptable service strategies. For example, it customizes personalized learning plans based on users' learning progress and knowledge acquisition; it intelligently allocates resources according to users' work tasks, achieving intelligent and personalized decision-making output and driving the precision of system services.

[0062] In one possible implementation of this application, generating service instructions based on user profile tags corresponding to dimensions with a weight percentage higher than a second preset threshold in the user's global feature vector, and the interaction scenario, may include: prioritizing user profile tags based on the user profile tags corresponding to dimensions with a weight percentage higher than the second preset threshold in the user's global feature vector, wherein the weight is proportional to the priority; obtaining the service type matching the user profile tags; and generating structured service instructions based on the service type and the interaction scenario, wherein the service instructions include at least one of the following: service objective, execution subject, operation steps, and timeliness requirements, and the service instructions carry a priority identifier.

[0063] In this embodiment, after determining the user profile tags corresponding to dimensions with a weight ratio higher than the second preset threshold, service priorities can be sorted according to the importance weight of the user profile tags. Service types that match high-weight tags are selected first. Then, service matching is performed in conjunction with a preset service rule base, such as feature-service mapping relationships and service execution constraints, to determine the service type. The service content details are then adjusted in conjunction with the user's current interaction context to generate structured service instructions. The instruction content can include service objectives, such as meeting a certain type of user need, execution subject, such as specifying an intelligent agent or terminal module, operation steps, such as the specific process of service execution, and timeliness requirements, such as the service completion time limit. Priority identifiers are added to the instructions. High-priority instructions are marked as immediate execution and must be started within a preset short time limit; medium- and low-priority instructions are marked as on-demand scheduling and can be flexibly arranged according to system resource usage.

[0064] This embodiment can prioritize services by tag weight and adjust service details based on user interaction scenarios, such as time and terminal type, to generate structured instructions. For example, in an education scenario, the system can prioritize pushing corresponding subject tutoring content based on the user's weak subject high-weight tags and the current learning time, and adapt to the display format of the terminal device. Compared with the unified standardized strategy used in related technologies, this can improve the accuracy of service matching and user satisfaction.

[0065] like Figure 2 The diagram illustrates a service strategy. Based on user profile tags and interaction scenarios, service instructions are generated, and different instructions are assigned priority levels. For example, handling product quality issues is marked as high priority, meaning it needs immediate attention; tracking logistics progress is marked as medium priority; and membership service benefits are marked as normal priority, meaning they can be handled last or not at all.

[0066] like Figure 3 The diagram shown is a flowchart of an interaction method provided in an embodiment of this application. Figure 3 As shown, the interaction method may include the content shown in steps 301 to 303.

[0067] In S301, the memory data of the intelligent agent and the user characteristic data of the application platform are acquired.

[0068] The memory data of the intelligent agent may include data such as user operation feedback, task execution trajectory, and interactive content annotation collected in real time during the execution of tasks by establishing real-time connections with various functional intelligent agents through APIs. It may also include data generated during the execution of tasks by other intelligent agents, subject to actual application, and this embodiment does not limit it.

[0069] User characteristic data may include user operation behavior, function usage frequency, device usage parameters, and other data that can reflect user characteristics, depending on the actual application; this embodiment does not impose any limitations. User operation behavior, function usage frequency, device usage parameters, and other characteristic data can be collected from various downstream application services through software development kits (SDKs) embedding or standardized interface calls.

[0070] This embodiment solves the problems of fragmented data collection and narrow coverage in related technologies by comprehensively collecting raw data.

[0071] In one example, real-time incremental acquisition can be used to ensure the timeliness of data, and the real-time incremental acquisition data can be categorized and labeled according to memory data or user characteristic data in order to distinguish the acquired data.

[0072] In S302, memory data and user characteristic data are sent to the cloud.

[0073] In S303, service instructions are received from the cloud. These service instructions are used to instruct the intelligent agent to perform tasks corresponding to the interaction scenario.

[0074] The interactive scenarios may include intelligent customer service scenarios, intelligent education service scenarios, intelligent office collaboration scenarios, etc., and may also include other interactive scenarios that intelligent agents can perform. This embodiment does not limit these scenarios and the actual application shall prevail.

[0075] In this embodiment, after receiving a service instruction from the cloud, the intelligent agent can perform corresponding tasks based on the service instruction, such as answering questions, resolving doubts, and querying information.

[0076] In one instance, service results can also be presented to users through multimodal display methods, such as text, voice, and images, adapting to the interaction habits of different users, ensuring that users can intuitively obtain personalized services, and realizing the implementation of intelligent services on the terminal side.

[0077] In the embodiments of this application, the memory data of the intelligent agent and the user characteristic data of the application platform are first acquired. Then, the memory data and user characteristic data are sent to the cloud. Finally, service instructions are received from the cloud, which are used to instruct the intelligent agent to perform tasks corresponding to the interaction scenario. The data acquired in the embodiments of this application includes not only the memory data of the intelligent agent but also the user characteristic data of the application platform. Compared with related technologies that only use intelligent agent data to serve users, the data is more diversified and more accurately portrays the user profile, thereby providing users with more personalized needs and improving service accuracy.

[0078] like Figure 4 The diagram shows a user interface. It illustrates how commands sent from the cloud can be translated into responses to user questions, providing a relatively satisfactory answer based on the user's specific needs.

[0079] In one possible implementation of this application, the interaction method may further include: acquiring incremental data, the incremental data including at least one of the following: incremental memory data of the agent, incremental user characteristic data of the application platform, and user feedback data; and sending the incremental data to the cloud.

[0080] This embodiment can send only the newly generated incremental data to the cloud when the intelligent agent or application platform generates new data or receives feedback data from the user, thereby reducing the amount of data transmission.

[0081] like Figure 5 The diagram illustrates the multi-layered architecture of this application. This embodiment employs a collaborative architecture comprising a user terminal layer, a network transmission layer, and a cloud engine layer. Two types of core data are collected from the user terminal layer, uploaded to the cloud engine layer via the network transmission layer, and preprocessed, deeply fused, updated, and generated as instructions. Then, instructions are issued via the network layer, and the terminal layer executes the service and collects feedback, forming a closed-loop end-to-end system. The user terminal layer includes an application platform interaction module and an intelligent agent interaction module, while the cloud engine layer includes a multi-source data preprocessing module, a memory-user feature fusion module, and an adaptive intelligent decision-making module.

[0082] Among them, the intelligent agent interaction module of the user terminal layer is used to collect the memory data of the intelligent agent, and the application platform interaction module is used to collect user feature data in downstream application services.

[0083] The network transport layer receives two types of data collected by the user terminal layer, encrypts them, and then uploads them to the cloud engine layer.

[0084] The multi-source data preprocessing module in the cloud engine layer preprocesses the received data to provide high-quality, standardized input data for subsequent data fusion, solving the problems of chaotic data formats and low validity in traditional systems. The memory and user feature data fusion module fuses multi-source data to obtain multi-dimensional feature vectors, which represent different user profile tag types, thereby determining user needs. The service instruction generation module prioritizes services according to the importance weight of user profile tags, and then, combined with the interaction scenario, issues service instructions to the network transport layer, which in turn issues the instructions to the user terminal layer.

[0085] In one example, the user terminal layer is the interface through which the entire system architecture interacts directly with the user, and it is also an important entry point for data collection, containing key interaction modules such as intelligent agent interaction and application platform interaction.

[0086] The intelligent agent interaction module enables bidirectional data interaction with the intelligent agent. On one hand, it issues task instructions to the intelligent agent, driving it to perform various tasks, such as answering questions and providing learning support in the education field, and information retrieval and transaction processing in the service field. On the other hand, during the intelligent agent's task execution, it collects the memory data generated by the intelligent agent in real time. This data records the process of interaction between the intelligent agent and the user, user feedback, and other information, providing raw materials for subsequent data processing and user profile construction.

[0087] The application platform interaction module interacts with downstream application services' applications (APPs) to collect user characteristic data. This data covers various aspects of user behavior on the APP, such as usage time and functional preferences, reflecting user characteristics and needs from different dimensions, enriching the user profile, and providing a more comprehensive basis for intelligent decision-making in the cloud. Simultaneously, this module receives policies from the network transport layer and executes these policies on the terminal side, presenting personalized services generated in the cloud to the user.

[0088] As the core decision-making and data processing hub of the entire system architecture, the cloud engine layer integrates modules such as multi-source data preprocessing, memory-user feature fusion, and adaptive intelligent decision-making to centrally manage, deeply integrate, and generate intelligent decisions for the entire system's data.

[0089] Multi-source data preprocessing module: This module unifies the processing of heterogeneous data from multiple sources, including agent interactions and application platform interactions. Agents generate memory data containing task execution trajectories and interaction content during task execution, while application platform interactions provide characteristic data such as user behavior habits and preference settings. These data often differ in format and standardization. This module cleanses the data, removing invalid and erroneous data; transforms the data into a unified format recognizable by the system; and reduces the data volume while retaining key information through data reduction, providing high-quality, standardized input for subsequent data fusion.

[0090] The memory-user feature fusion module establishes a mapping relationship between agent memory data and user feature data based on preprocessed data. It can mine information such as user needs and behavioral patterns reflected in the agent's memory, and deeply integrate this information with basic user characteristics and historical behavior data collected from the application platform to construct a more comprehensive and three-dimensional user profile. This solves the problems of data isolation and difficulty in fully characterizing users in traditional systems, providing a data foundation for precise services.

[0091] The adaptive intelligent decision-making module, based on integrated user profiles and multi-source data, combined with pre-set decision models and algorithms, can autonomously adjust decision logic according to different application scenarios and user needs, generating highly adaptable service strategies. For example, it can customize personalized learning plans based on users' learning progress and knowledge acquisition; and intelligently allocate resources according to users' work tasks, achieving intelligent and personalized decision output and driving the precision of system services.

[0092] The network transport layer serves as a bridge connecting the cloud engine layer and the user terminal layer, ensuring efficient and stable transmission of data and policies between these layers. It supports the reliable uploading of agent-memory data and user feature data from the user terminal layer to the cloud engine layer, employing technologies such as encrypted transmission and data fragmentation to ensure data security and integrity during transmission, preventing data leakage or corruption. Simultaneously, it can accurately distribute policies generated by the cloud engine layer to the user terminal layer, providing a channel for policy execution on the terminal and ensuring that cloud decisions are delivered to the terminal side in a timely and accurate manner, achieving collaborative linkage between the cloud and the terminal.

[0093] In one instance, after the network transport layer receives two types of data collected by the user terminal layer, it initiates an encrypted upload process.

[0094] First, the data can be encrypted using a compliant symmetric encryption algorithm to generate encrypted data packets with unique identifiers. These unique identifiers include the data source terminal ID, collection timestamp, and data type marker, facilitating cloud-based traceability and classification. Then, a secure end-to-cloud channel is established based on a secure transmission protocol, such as the Secure Sockets Layer / Transport Layer Security Protocol (SSL / TLS). Transmission links are allocated according to the real-time requirements of the data—high-real-time intelligent agent memory data prioritizes low-latency network links, while batch application characteristic data employs a staggered transmission strategy, such as transmitting during idle periods of terminal devices, to avoid network resource conflicts. Finally, the encrypted data packets are reliably uploaded to the cloud engine layer, supporting breakpoint resumption. If transmission is interrupted, it can resume from the breakpoint without re-uploading the entire dataset, ensuring the security, continuity, and efficiency of data transmission.

[0095] After receiving personalized service instructions generated by the cloud engine layer, the network transport layer classifies and encapsulates the instructions to distinguish between intelligent agent execution instructions and terminal display instructions. It clarifies the target execution subject, priority, and timeliness requirements of the instructions, transmits the intelligent agent execution instructions to the intelligent agent interaction module of the user terminal layer through a dedicated communication protocol, and transmits the terminal display instructions to the application platform interaction module of the user terminal layer through an encrypted channel. This ensures that cloud decisions can be transmitted to the terminal side in a timely and accurate manner, realizes the collaborative linkage between the cloud and the terminal, and avoids service disconnection caused by instruction transmission delays or loss.

[0096] After collecting incremental data at the user terminal layer, the network transmission layer uploads the encrypted feedback data to the cloud engine layer through a secure channel, providing a basis for subsequent end-to-end optimization in the cloud, such as adjusting data fusion weights and service matching rules.

[0097] In one embodiment of this application, the interaction scenario is an intelligent customer service scenario, which can cover fields such as e-commerce, finance, and lifestyle services. It needs to achieve the goals of rapid response to user needs and accurate location of service problems. Based on the above embodiments of this application, this scenario includes the following steps: Data collection is conducted at the user terminal layer: User terminals, such as mobile apps and computer web pages, connect to customer service AI agents through the AI ​​agent interaction module to collect interaction memory data in real time, such as dialogue records of user complaints about damaged goods not being replaced, and feedback content on inquiries about refund processing time. At the same time, the application platform interaction module connects to enterprise service systems, such as e-commerce order systems and financial account systems, to collect user characteristic data, such as the number of return and exchange orders in recent months, such as the number of historical refund application records, and the types of services frequently consulted in the past three months. User privacy data, such as payment card numbers and addresses, are uploaded to the cloud engine layer after being anonymized.

[0098] Data fusion is performed at the cloud engine layer: The cloud first constructs a dynamic knowledge graph containing agent memory nodes, feature attribute nodes, etc., and fills the agent memory nodes with memory data such as product damage complaints and refund timeliness inquiries. The agent memory nodes can contain attributes such as memory type and memory content summary, and are associated with user nodes through memory ownership edges; feature data such as high-frequency returns and exchanges and no historical refund records are filled with feature attribute nodes, and are associated with user nodes through feature ownership edges. Then, a GNN model is deployed to aggregate the features of two types of nodes to generate embedding vectors: The vector similarity between the product damage complaint memory node and the high-frequency return / exchange feature node is calculated to determine if it falls within a threshold range. For example, if the vector similarity between the product damage complaint memory node and the high-frequency return / exchange feature node is 0.5, it falls within the threshold range, triggering conflict reconciliation. The GNN then traces back the interaction scenario node, labeling the current issue as a quality problem and past issues as logistics problems. Combining the credibility of the data source node (e.g., if the credibility of order system data is higher than historical records), the node association weights are adjusted, and a product quality after-sales demand label is generated. The similarity between the refund timeliness inquiry memory node and the feature node with no historical refund record is calculated to see if it is less than or equal to the minimum threshold. If so, the GNN determines there is no association and automatically generates a new first-time refund inquiry label for both types of nodes. Finally, the after-sales complaint label is weighted higher than the inquiry label, completing the user service demand profile construction.

[0099] Decision-making and Service: The cloud-based system incrementally updates the user service demand model based on user profiles, determining that the user's current core needs are product reshipment and refund timeliness confirmation. The adaptive decision-making module assigns high-weight tags based on product quality and after-sales needs, combined with the user's current online status and urgent need for feedback, to generate service instructions that prioritize transferring the user to a dedicated after-sales customer service representative and push notifications for reshipment progress inquiries. The network transmission layer can issue instructions via low-latency links, such as 5G / Wi-Fi 6, and the terminal simultaneously presents the information through a combination of text pop-ups and voice prompts. After the service is provided, user satisfaction feedback on problem resolution is collected, and the cloud adjusts the associated weights of corresponding nodes in the graph accordingly. If the satisfaction level is high, the weight of the node associated with the dedicated customer service tag is increased, forming a closed loop.

[0100] In another embodiment of this application, the interaction scenario is an intelligent education service scenario. The difference between this scenario and the intelligent customer service scenario is that this scenario focuses on personalized learning guidance and dynamic progress adaptation.

[0101] Data collection is performed at the user terminal layer: student terminals, such as learning tablets and mobile apps, connect to the tutoring agent through the intelligent agent interaction module to collect interactive memory data, such as dialogue records of asking about the drawing of geometric auxiliary lines, and memory data such as feedback on the correction of wrong questions; and connect to the education app through the application platform interaction module to collect feature data, such as the error rate of function questions and the learning time of geometric knowledge points.

[0102] Data fusion is performed at the cloud engine layer: A dynamic knowledge graph is constructed in the cloud, and the memory data of geometric auxiliary line queries is filled into the agent's memory nodes, while the error rate of function questions is filled into the feature attribute nodes; a GNN model is deployed to aggregate node features, and the similarity between the geometric auxiliary line query memory node and the geometric learning time feature node is calculated to see if it is greater than or equal to the highest threshold. If so, the GNN strengthens the association between the two and merges them to generate a geometric weakness label; at the same time, based on the function question error rate feature node, a new label for function reinforcement is generated, and the label weights are adjusted according to the importance of knowledge points, such as the weight of geometric weakness being higher than that of function reinforcement, thus constructing a learning needs profile.

[0103] Decision-making and services: The cloud updates the learning model based on profiles, the decision module pushes auxiliary line teaching videos and geometry-specific exercises, and the terminal displays them in a combination of video and text. After the service, feedback on the accuracy of the exercises is collected, the cloud updates the graph node attributes, and if the accuracy is low, it supplements the memory node attributes of geometric auxiliary lines and optimizes the label weights.

[0104] In another embodiment of this application, the interaction scenario is an intelligent office collaboration scenario. The difference between this scenario and the intelligent customer service scenario mentioned above is that this scenario focuses on task collaboration and dynamic resource allocation.

[0105] Data collection is conducted at the user terminal layer: employee terminals, such as office computers and enterprise apps, connect to task agents through the intelligent agent interaction module to collect interactive memory data, such as memory data of insufficient permissions for project documents and task progress report records; and connect to the office system through the application platform interaction module to collect feature data, such as the frequency of project participation and document editing records in the past week.

[0106] Data fusion is performed at the cloud engine layer: a graph is constructed in the cloud, document permission feedback memory data is filled into the agent's memory nodes, and project participation frequency is filled into the feature attribute nodes; after the GNN model aggregates features, it calculates whether the similarity between the permission feedback memory nodes and the high participation frequency feature nodes is within a threshold range. If so, the GNN backtracks to the project scene nodes, marks the core task participation, reconciles the correlation logic, and merges to generate project core member tags; at the same time, based on the document editing record feature nodes, high-frequency editor tags are supplemented, and the weights are adjusted according to task priority.

[0107] Decision-making and services: The collaborative model is updated in the cloud, and the decision-making module issues instructions to grant open document editing permissions, which are automatically applied in the terminal office software; feedback on the ease of use of permissions is collected after the service is provided, and the graph edge association weights are updated in the cloud. If the ease of use is high, the weights of the core members and document permission edges are enhanced, and the resource allocation logic is optimized.

[0108] Figure 6 This is a schematic diagram of an interactive device provided in an embodiment of this application. Figure 6As shown, the interactive device may include a first receiving module 601, a fusion module 602, and a generation module 603.

[0109] The first receiving module 601 is used to receive multi-source data, including the agent's memory data and the application platform's user feature data; the fusion module 602 is used to perform feature fusion processing on the multi-source data to obtain a user global feature vector, the dimension of which represents the user profile label type; the generation module 603 is used to generate service instructions based on the dimension in the user global feature vector and the interaction scenario, the service instructions being used to instruct the agent to perform a task corresponding to the interaction scenario.

[0110] In the embodiments of this application, the first receiving module 601 first receives multi-source data, including the agent's memory data and the application platform's user feature data. Then, the fusion module 602 performs feature fusion processing on the multi-source data to obtain a global user feature vector. The dimension of the global feature vector represents the user profile label type. Finally, the generation module 603 generates a service instruction based on the dimension in the global user feature vector and the interaction scenario. The service instruction is used to instruct the agent to perform a task corresponding to the interaction scenario. This embodiment of the application can fuse the received multi-source data. Compared with related technologies that only use agent data to serve users, the data is more diversified. The fused global user feature vector makes the profile dimensions diverse. Combined with the scenario, it can provide users with more personalized needs and improve service accuracy.

[0111] In one possible implementation of this application, the fusion module 602 is used to: construct a knowledge graph based on multi-source data; and perform feature fusion on the knowledge graph based on a graph neural network model to obtain a user global feature vector.

[0112] In one possible implementation of this application, the fusion module 602 is used to: preprocess the memory data and user feature data to obtain standardized data, wherein the preprocessing includes at least one of the following: data cleaning, format conversion, and data reduction.

[0113] In one possible implementation of this application, the fusion module 602 is used to: determine nodes in the knowledge graph based on memory data and user feature data, wherein the nodes include preset attributes; and fill the agent's memory data and user feature data into each node in the knowledge graph according to the preset attributes and correspondences to obtain the knowledge graph.

[0114] In one possible implementation of this application, the fusion module 602 is used to: determine a graph neural network model based on the number of nodes and edge density of the knowledge graph; convert the data in the knowledge graph into a feature matrix, the feature matrix including node features; aggregate and propagate the feature matrix using the graph neural network model to obtain a target knowledge graph, the target knowledge graph including the node attributes and edge weights of each node; and perform feature fusion on the target knowledge graph using the graph neural network model to obtain a user global feature vector.

[0115] In one possible implementation of this application, the fusion module 602 is used to: input node features into a graph neural network model for aggregation and propagation to obtain bidirectional features between multiple nodes; generate low-dimensional dense embedding vectors for each node using the graph neural network model based on the feature aggregation results; calculate the association strength between any two nodes based on a preset algorithm; and obtain the target knowledge graph based on the bidirectional features between multiple nodes, the low-dimensional dense embedding vectors of each node, and the association strength between any two nodes.

[0116] In one possible implementation of this application, the fusion module 602 is configured to: determine that there is a conflict between nodes when nodes meet a first preset condition, the first preset condition including at least one of the following: the association strength between any two nodes is less than a first preset threshold, the same node has different node attributes, and the data between different nodes contradict each other; when there is a conflict between nodes, analyze the cause of the conflict based on the association path of the conflicting nodes and obtain the attribution result; and adjust the attributes of the conflicting nodes based on the attribution result to obtain the updated node attributes and edge weights.

[0117] In one possible implementation of this application, the fusion module 602 is configured to: update the knowledge graph based on the received data when the received data meets a second preset condition, wherein the second preset condition includes at least one of the following: new memory data uploaded by the terminal, new user feature data, or user feedback data.

[0118] In one possible implementation of this application, the generation module 603 is used to: obtain user profile tags corresponding to dimensions whose weight percentage in the user's global feature vector is higher than a second preset threshold; and generate service instructions based on the user profile tags corresponding to dimensions whose weight percentage in the user's global feature vector is higher than the second preset threshold and the interaction scenario.

[0119] In one possible implementation of this application, the generation module 603 is used to: prioritize user profile tags based on user profile tags corresponding to dimensions with a weight ratio higher than a second preset threshold in the user's global feature vector, wherein the weight is proportional to the priority; obtain the service type matching the user profile tags; and generate structured service instructions based on the service type and interaction scenario, wherein the service instructions include at least one of the following: service objective, execution subject, operation steps, and timeliness requirements, and the service instructions carry a priority identifier.

[0120] The interactive device provided in this application embodiment can achieve... Figures 1 to 2 The various processes implemented in the method embodiments achieve the same technical effect, and will not be described again here to avoid repetition.

[0121] Figure 7 This is a schematic diagram of an interactive device provided in an embodiment of this application. Figure 7 As shown, the interactive device may include an acquisition module 701, a sending module 702, and a second receiving module 703.

[0122] The acquisition module 701 is used to acquire the memory data of the intelligent agent and the user feature data of the application platform; the sending module 702 is used to send the memory data and user feature data to the cloud; and the second receiving module 703 is used to receive the service instructions sent from the cloud, which are used to instruct the intelligent agent to perform tasks corresponding to the interaction scenario.

[0123] In the embodiments of this application, the acquisition module 701 first acquires the memory data of the intelligent agent and the user characteristic data of the application platform. Then, the sending module 702 sends the memory data and user characteristic data to the cloud. Finally, the second receiving module 703 receives the service instructions sent from the cloud. The service instructions are used to instruct the intelligent agent to perform tasks corresponding to the interaction scenario. The data acquired in the embodiments of this application includes not only the memory data of the intelligent agent but also the user characteristic data of the application platform. Compared with related technologies that only use intelligent agent data to serve users, the data is more diversified and more accurately portrays the user profile, thereby providing users with more personalized needs and improving service accuracy.

[0124] In one possible embodiment of this application, the interactive device may further include: a second acquisition module and a second transmission module.

[0125] The second acquisition module is used to acquire incremental data, which includes at least one of the following: incremental memory data of the intelligent agent, incremental user feature data of the application platform, and user feedback data; the second sending module is used to send the incremental data to the cloud.

[0126] The interactive device provided in this application embodiment can achieve... Figures 3 to 4The various processes implemented in the method embodiments achieve the same technical effect, and will not be described again here to avoid repetition.

[0127] like Figure 8 As shown, this application embodiment also provides a communication device 800, including a processor 801, a memory 802, and a program or instructions stored in the memory 802 and executable on the processor 801. When the program or instructions are executed by the processor 801, they implement the various processes of the above-described interaction method embodiments and achieve the same technical effects. To avoid repetition, they will not be described again here.

[0128] This application also provides a storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the interactive method embodiments provided in any of the above embodiments. Furthermore, they achieve the same technical effects, and to avoid repetition, will not be described again here.

[0129] The processor is the processor in the communication device described in the above embodiments. The storage medium includes computer storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0130] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described interaction method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0131] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0132] This application also provides a computer program / program product, which is stored in a storage medium and executed by at least one processor to implement the various processes of the above-described interactive method embodiments, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0133] This application embodiment also provides a processing device, which is configured to execute the various processes of the above-described interactive method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0134] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0135] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0136] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An interaction method, characterized in that, include: Receive multi-source data, including the memory data of the intelligent agent and the user feature data of the application platform. The memory data of the intelligent agent includes user operation feedback, task execution trajectory, and interactive content annotation. The user feature data includes user operation behavior, function usage frequency, and device usage parameters. The multi-source data is subjected to feature fusion processing to obtain a global user feature vector, and the dimension of the global user feature vector represents the user profile label type. Based on the dimensions in the user's global feature vector and the interaction scenario, a service instruction is generated. The service instruction is used to instruct the agent to perform a task corresponding to the interaction scenario. The process of generating service instructions based on the dimensions in the user's global feature vector and the interaction scenario includes: Obtain the user profile labels corresponding to the dimensions whose weight percentage in the user's global feature vector is higher than the second preset threshold; Based on the user profile tags corresponding to the dimensions with a weight ratio higher than the second preset threshold in the user's global feature vector, and the interaction scenario, service instructions are generated.

2. The method according to claim 1, characterized in that, The step of performing feature fusion processing on the multi-source data to obtain the user's global feature vector includes: Based on the aforementioned multi-source data, a knowledge graph is constructed; The knowledge graph is fused using a graph neural network model to obtain a global feature vector for the user.

3. The method according to claim 2, characterized in that, Before constructing the knowledge graph based on the multi-source data, the method includes: The memory data and the user feature data are preprocessed to obtain standardized data. The preprocessing includes at least one of the following: data cleaning, format conversion, and data reduction.

4. The method according to claim 2, characterized in that, The construction of a knowledge graph based on the multi-source data includes: Based on the memory data and the user feature data, nodes in the knowledge graph are determined, and the nodes include preset attributes; The agent's memory data and the user's feature data are filled into each node of the knowledge graph according to the preset attributes and correspondences to obtain the knowledge graph.

5. The method according to claim 2, characterized in that, The feature fusion of the knowledge graph based on the graph neural network model to obtain the user's global feature vector includes: Based on the number of nodes and edge density of the knowledge graph, a graph neural network model is determined; The data in the knowledge graph is transformed into a feature matrix, which includes node features; Using the graph neural network model, the feature matrix is ​​aggregated and propagated to obtain a target knowledge graph, which includes the node attributes and edge weights of each node. The target knowledge graph is fused using the graph neural network model to obtain a global feature vector for the user.

6. The method according to claim 5, characterized in that, The process of using the graph neural network model to aggregate and propagate the feature matrix to obtain the target knowledge graph includes: The node features are input into the graph neural network model for aggregation and propagation to obtain bidirectional features between multiple nodes. Based on the feature aggregation results, the graph neural network model is used to generate low-dimensional dense embedding vectors for each node. The association strength between any two nodes is calculated based on a preset algorithm; Based on the bidirectional features between multiple nodes, the low-dimensional dense embedding vectors of each node, and the correlation strength between any two nodes, the target knowledge graph is obtained.

7. The method according to claim 6, characterized in that, The method further includes: If a node meets a first preset condition, it is determined that there is a conflict between the nodes. The first preset condition includes at least one of the following: the association strength between any two nodes is less than a first preset threshold, the same node has different node attributes, and the data between different nodes contradict each other. In the case of conflicts between nodes, the causes of the conflicts are analyzed based on the association paths of the conflicting nodes, and the attribution results are obtained. Based on the attribution results, the attributes of the conflicting nodes are adjusted to obtain the updated node attributes and edge weights.

8. The method according to claim 2, characterized in that, The method further includes: If the received data meets the second preset condition, the knowledge graph is updated based on the received data. The second preset condition includes at least one of the following: new memory data uploaded by the terminal, new user feature data, and user feedback data.

9. The method according to claim 1, characterized in that, The step of generating service instructions based on user profile tags corresponding to dimensions with weight percentages higher than a second preset threshold in the user's global feature vector, and the interaction scenario, includes: Based on the user profile tags corresponding to the dimensions with a weight ratio higher than the second preset threshold in the user global feature vector, the user profile tags are prioritized, wherein the weight is proportional to the priority. Obtain the service type matched by the user profile tags; Based on the service type and interaction scenario, a structured service instruction is generated. The service instruction includes at least one of the following: service objective, execution subject, operation steps, and timeliness requirements. The service instruction carries a priority identifier.

10. An interaction method, characterized in that, include: Acquire the memory data of the intelligent agent and the user feature data of the application platform. The memory data of the intelligent agent includes user operation feedback, task execution trajectory, and interactive content annotation. The user feature data includes user operation behavior, function usage frequency, and device usage parameters. Send the memory data and the user characteristic data to the cloud; The agent receives service instructions from the cloud. The service instructions are used to instruct the agent to perform tasks corresponding to the interaction scenario. The service instructions are generated based on the dimensions in the user's global feature vector and the interaction scenario. The generation process of the service instructions includes: obtaining user profile tags corresponding to the dimensions in the user's global feature vector whose weight ratio is higher than a second preset threshold. Based on the user profile tags corresponding to the dimensions with a weight ratio higher than the second preset threshold in the user's global feature vector, and the interaction scenario, service instructions are generated.

11. The method according to claim 10, characterized in that, The method further includes: Acquire incremental data, which includes at least one of the following: incremental memory data of the intelligent agent, incremental user feature data of the application platform, and user feedback data; The incremental data is sent to the cloud.

12. A communication device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as claimed in any one of claims 1 to 9, or implement the steps of the method as claimed in claim 10 or 11.

13. A computer program product, characterized in that, The computer program product is stored in a storage medium and is executed by at least one processor to implement the steps of the method as claimed in any one of claims 1 to 9, or to implement the steps of the method as claimed in claim 10 or 11.

Images