Multi-dimensional fusion meta universe and vertical AI model collaborative innovation platform

Abstract

The invention relates to the technical field of meta universe and knowledge graph, and discloses a multi-dimensional fusion meta universe and vertical AI model collaborative innovation platform, which comprises a data generation module, an interaction intention knowledge graph construction module, a vertical domain knowledge graph construction module, a knowledge graph fusion module and an intelligent interaction module. The method comprises the following steps: collecting user interaction behavior data, identifying dominant and implicit intentions, and constructing an interaction intention knowledge graph; generating a fused heterogeneous knowledge graph in combination with the vertical domain knowledge graph; constructing a vertical AI model based on the fused heterogeneous knowledge graph; according to the invention, deep understanding and prediction of user intentions are realized, so that intelligent personalized services are provided.

Description

Technical Field

[0001] The present invention relates to the field of metaverse and knowledge graph technology, and more specifically, to a multi-dimensionally integrated metaverse and vertical AI model collaborative innovation platform. Background Art

[0002] With the rapid development of the concept and technology of the metaverse, human-computer interaction and AI digital humans have received extensive attention. There are some patents related to human-computer interaction of intelligent devices and building AI digital humans based on AIGC in the prior art, but there are still some shortcomings.

[0003] A Chinese patent with publication number CN117540017A discloses a method, apparatus and smart device for human-computer interaction of smart devices. The method obtains the total text data of the perception data in the multi-source heterogeneous data of human-computer interaction of smart devices, generates perception interaction data based on the associated prompt information, and classifies the perception interaction data by scene application, thereby improving the accuracy of human-computer interaction of smart devices. However, the patent mainly focuses on using text data for human-computer interaction, and has limited ability to understand and predict user intentions, and cannot provide more intelligent and personalized services.

[0004] The Chinese patent with publication number CN117251057A discloses a method and system for building an AI digital person based on AIGC. The method builds an AI digital person by determining needs and goals, making an AI digital person, performing data collection and preprocessing, feature engineering and selection, model selection and training, system development and integration, etc. AI digital people can collect and analyze a large amount of customer data, practitioner data, and industry data to provide support for industry decision-making and improve operational efficiency. However, the AI ​​digital person constructed in this patent is mainly based on single-dimensional data and models, lacks in-depth understanding and integration of user interaction intentions and vertical field knowledge, and is difficult to adapt to the complex and changeable metaverse application scenarios.

[0005] To sum up, although the existing technology has made certain progress in the human-computer interaction of smart devices and the construction of AI digital humans, the understanding of user intentions is not deep enough, the utilization of vertical field knowledge is insufficient, and there is a lack of multi-dimensional integration, which makes it impossible to effectively support intelligent and personalized services in the metaverse environment. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-dimensional integrated metaverse and vertical AI model collaborative innovation platform, which collects user interaction behavior data to identify intentions and constructs an interaction intention knowledge graph, combines the vertical field knowledge graph for heterogeneous fusion, and builds a vertical AI model on this basis, thereby achieving a deep understanding and prediction of user intentions and providing intelligent personalized services.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform, including: data generation module, interaction intention knowledge graph construction module, vertical field knowledge graph construction module, knowledge graph fusion module and intelligent interaction module;

[0009] The data generation module is used to collect the user's interactive behavior data, extract a first data set from the interactive behavior data; analyze the first data set, identify the user's explicit intention and implicit intention, and generate a second data set;

[0010] The interaction intention knowledge graph construction module is used to construct a user-centric interaction intention knowledge graph based on the second data set;

[0011] The vertical domain knowledge graph construction module is used to identify key vertical domain knowledge for different metaverse application scenarios and construct a vertical domain knowledge graph based on the vertical domain knowledge;

[0012] The knowledge graph fusion module is used to fuse the vertical domain knowledge graph with the interactive intention knowledge graph to obtain a fused heterogeneous knowledge graph;

[0013] The intelligent interaction module is used to build a vertical AI model based on the fused heterogeneous knowledge graph, apply the vertical AI model to understand and predict user intentions, and realize intelligent personalized services.

[0014] Further, the data generation module includes a first data generation unit and a second data generation unit;

[0015] The first data generation unit is used to record the user's interactive operations including clicks, slides and voice inputs in real time to form an original interaction log; extract key interaction fields including user ID, interaction object ID, interaction timestamp and interaction type from the original interaction log to form a structured interaction behavior record; aggregate the interaction behavior records according to user ID to generate a first data set, which includes the user's interaction behavior sequence.

[0016] Furthermore, the second data generation unit is used to encode the interaction behavior sequence in the first data set, extract the semantic feature vector of the interaction behavior sequence; cluster the semantic feature vector, identify the explicit intention according to the clustering result, and associate it with the predefined intention type to obtain the explicit intention type; establish an intention evolution model, obtain the hidden state sequence of the user interaction behavior sequence through the intention evolution model, and obtain the user's implicit intention and implicit intention type according to the hidden state sequence; organize the identified explicit intention and implicit intention in a hierarchical manner, and generate a second data set in combination with the time sequence of the intention occurrence, wherein each data sample in the second data set includes an intention type, an intention hierarchy, and an intention time sequence;

[0017] The second data generating unit includes a data encoding unit, an explicit intention recognition unit, an implicit intention recognition unit and a data combining unit.

[0018] Furthermore, the data encoding unit is used to map each interaction event in the interaction behavior sequence into a low-dimensional dense vector as an embedded representation of the interaction event, wherein the interaction behavior sequence includes n interaction events; combine the embedded representations of the n interaction events into an event embedding sequence; and obtain a semantic feature vector of the interaction behavior sequence according to the event embedding sequence;

[0019] The explicit intention recognition unit is used to calculate the similarity matrix between semantic feature vectors; construct a minimum spanning tree based on the similarity matrix; and set a distance threshold on the minimum spanning tree. , according to the distance threshold Obtaining a cluster of semantic feature vectors; matching each cluster of semantic feature vectors with a predefined intent type to obtain an explicit intent type of the cluster;

[0020] The implicit intention recognition unit is used to establish an intention evolution model, obtain a hidden state sequence of the user interaction behavior sequence through the intention evolution model, and obtain the user's implicit intention and implicit intention type according to the hidden state sequence;

[0021] The data combination unit is used to construct a hierarchical intent type system based on the semantic relationship between intent types; map each user interaction behavior sequence to the intent type system, and assign a corresponding intent type label to each interaction event in the interaction behavior sequence; record the timestamp of each intent type to obtain the time sequence of the intent sequence; and generate a second data set based on the time sequence of the intent sequence.

[0022] Furthermore, the interactive intention knowledge graph construction module includes a graph mapping unit, a node association unit and an attribute attachment unit;

[0023] The graph mapping unit is used to map the intent type, intent hierarchy, and intent sequence in the second data set into nodes and edges in the knowledge graph;

[0024] The node association unit is used to associate a series of intent instance nodes of the same user at different time points with the user node of the user, forming a star topology subgraph with the user node as the center and the intent instance nodes surrounding it;

[0025] The attribute adding unit is used to add attribute information to the nodes and edges in the knowledge graph to obtain the final interaction intention knowledge graph.

[0026] Furthermore, the mapping of intent types, intent hierarchies, and intent sequences in the second data set to nodes and edges in the knowledge graph includes: mapping intent types to intent type nodes in the knowledge graph; mapping intent instances at different time points in the intent sequence in the second data set to intent instance nodes under corresponding type nodes; constructing intent sequence edges between intent instance nodes based on intent sequence; mapping each user to a user node in the knowledge graph; establishing subclass relationship edges between corresponding intent type nodes based on the hierarchical relationship of intent types; and establishing instance relationship edges between intent instance nodes and corresponding intent type nodes.

[0027] Furthermore, the vertical field knowledge graph construction module includes a knowledge identification unit and a graph construction unit;

[0028] The knowledge identification unit is used to determine the key vertical fields of the metaverse application scenarios; for each vertical field, it collects structured knowledge including domain ontology, knowledge base and industry dictionary and unstructured knowledge including domain documents and user comments; and constructs vertical domain knowledge from structured knowledge and unstructured knowledge.

[0029] Furthermore, the knowledge graph fusion module includes an embedding vector generation unit, an embedding vector optimization unit and a graph fusion unit;

[0030] The embedding vector generating unit is used to learn the dense embedding representation of nodes and edges in the vertical domain knowledge graph using a graph neural network, and generate an initial graph embedding vector including a node embedding vector and an edge embedding vector; the node embedding vector includes an entity node embedding vector;

[0031] The embedding vector optimization unit is used to propagate the node embedding vectors in the initial graph embedding vector through the graph convolution network, so that entities with similar semantics or structures in the graph obtain similar node embedding vectors, and use the similar node embedding vectors to update the node embedding vectors in the initial graph embedding vector to obtain an optimized graph embedding vector;

[0032] The graph fusion unit is used to associate the optimized graph embedding vector of the vertical field with the nodes in the interactive intent knowledge graph to obtain a fused heterogeneous knowledge graph; the nodes in the interactive intent knowledge graph include user nodes and intent instance nodes.

[0033] Furthermore, the method of associating the optimized graph embedding vector of the vertical field with the nodes in the interactive intent knowledge graph includes: obtaining the node embedding vector of the interactive intent knowledge graph; calculating the similarity between the node embedding vector of the interactive intent knowledge graph and the node embedding vector of the vertical field knowledge graph, and calculating the similarity of the node embedding vector greater than the similarity threshold. The nodes of the interactive intention knowledge graph and the nodes of the vertical domain knowledge graph establish semantic associations between cross-graph nodes; the semantic associations between cross-graph nodes are used to construct a node alignment loss function; according to the node alignment loss function and the structure preservation loss of a single graph, a comprehensive optimization function is established, and the fused heterogeneous knowledge graph is obtained by minimizing the comprehensive loss function; the structure preservation loss of the single graph includes the structure preservation loss function of the interactive intention knowledge graph and the structure preservation loss function of the vertical domain knowledge graph.

[0034] A multi-dimensional fusion metaverse and vertical AI model collaborative innovation method, based on the multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform, the method comprises:

[0035] Collecting user interaction behavior data, extracting a first data set from the interaction behavior data; analyzing the first data set, identifying the user's explicit intention and implicit intention, and generating a second data set; constructing a user-centered interaction intention knowledge graph based on the second data set;

[0036] Identify key vertical domain knowledge for different metaverse application scenarios, build vertical domain knowledge graphs based on vertical domain knowledge, and fuse the vertical domain knowledge graphs with the interaction intent knowledge graphs to obtain fused heterogeneous knowledge graphs;

[0037] Based on the fused heterogeneous knowledge graph, a vertical AI model is constructed, and the vertical AI model is applied to understand and predict user intentions to achieve intelligent personalized services.

[0038] An electronic device includes a memory, a central processing unit, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned multi-dimensional fusion metaverse and vertical AI model collaborative innovation method when executing the computer program.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed, implements the above-mentioned multi-dimensional fusion metaverse and vertical AI model collaborative innovation method.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] By collecting and analyzing user interaction behavior data, the platform can identify users' explicit and implicit intentions and build a detailed user intention knowledge graph, which can provide a deeper understanding of user needs and improve user experience; based on the fused heterogeneous knowledge graph, the platform can build a vertical AI model and accurately predict user intentions, which enables the platform to provide highly personalized services to meet users' diverse needs; by integrating the vertical domain knowledge graph with the interaction intention knowledge graph, the platform can share and reuse knowledge in different application scenarios, enhancing the intelligence level of the system; by providing personalized and intelligent services, the platform can improve user satisfaction and stickiness, and increase user residence time and engagement in the metaverse environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 A module connection relationship diagram of a multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform in Example 1 of the present invention;

[0044] Figure 2 This is a flow chart of a multi-dimensional fusion metaverse and vertical AI model collaborative innovation method in Example 2 of the present invention;

[0045] Figure 3 A flowchart of a method for extracting a first data set from interactive behavior data in a multi-dimensional fusion metaverse and vertical AI model collaborative innovation method of the present invention;

[0046] Figure 4 A flowchart of a method for generating a second data set in a multi-dimensional fusion metaverse and vertical AI model collaborative innovation method of the present invention;

[0047] Figure 5 A schematic diagram of the structure of an electronic device provided in one embodiment of the present application;

[0048] Figure 6A schematic diagram of the computer-readable storage medium structure provided for one embodiment of the present application. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] Example 1

[0051] See also Figure 1 As shown, this embodiment provides a multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform, including: a data generation module, an interaction intention knowledge graph construction module, a vertical field knowledge graph construction module, a knowledge graph fusion module and an intelligent interaction module;

[0052] The data generation module is used to collect the user's interactive behavior data, extract a first data set from the interactive behavior data; analyze the first data set, identify the user's explicit intention and implicit intention, and generate a second data set;

[0053] The data generation module includes a first data generation unit and a second data generation unit;

[0054] The first data generation unit is used to record the user's interactive operations including clicks, slides and voice inputs in real time to form an original interaction log; extract key interaction fields including user ID, interaction object ID, interaction timestamp and interaction type from the original interaction log to form a structured interaction behavior record; aggregate the interaction behavior records according to user ID to generate a first data set, which includes the user's interaction behavior sequence.

[0055] Specifically, the first data generation unit comprehensively collects various interactive behaviors of users from the client and server of the Metaverse platform. Collect and record the real-time operations of users in 3D scenes, such as object selection, perspective change, gesture interaction, etc. from the client; collect and record the user's identity authentication, scene access, social interaction and other behaviors from the server; the original interaction log uses semi-structured data formats such as JSON to facilitate subsequent parsing and processing; the original interaction log is cleaned and filtered to remove redundant and noisy data, and key fields such as user ID, interaction object ID, interaction timestamp, interaction type, etc. are extracted to form a structured interactive behavior record; in the data cleaning stage, unqualified log records are eliminated through field integrity verification, timestamp anomaly detection and other methods, and unstructured data is converted into a structured field form. The first data set finally generated is organized in the form of a relational table or document database, recording the interaction behavior sequence of each user at different times and in different scenarios, providing a comprehensive and accurate data foundation for subsequent user intention analysis. The first data generation unit is used to collect and preprocess user interaction logs, providing key data support for user intent analysis; through collaborative collection between the client and the server, the first data generation unit covers various forms of interaction in the metaverse environment and comprehensively records the user's behavior trajectory. The cleaning and structured conversion of the original logs ensure the quality and consistency of the data and reduce the difficulty of subsequent processing. Aggregating the interaction records by user and abstracting a unified serialized representation is conducive to characterizing the continuity and context dependency of user behavior, laying the foundation for mining the evolution of user intent. At the same time, the standardized data organization form also facilitates the association and fusion with user portrait data from other sources, enriching the semantic representation of user behavior from multiple dimensions.

[0056] The second data generation unit is used to encode the interaction behavior sequence in the first data set and extract the semantic feature vector of the interaction behavior sequence; cluster the semantic feature vector, identify the explicit intention according to the clustering result, and associate it with the predefined intention type to obtain the explicit intention type; establish an intention evolution model, obtain the hidden state sequence of the user interaction behavior sequence through the intention evolution model, and obtain the user's implicit intention and implicit intention type according to the hidden state sequence; organize the identified explicit intentions and implicit intentions in a hierarchical manner, and generate a second data set in combination with the time sequence of the intention occurrence, wherein each data sample in the second data set includes an intention type, an intention hierarchy, and an intention sequence.

[0057] The second data generating unit includes a data encoding unit, an explicit intention recognition unit, an implicit intention recognition unit and a data combining unit;

[0058] The data encoding unit is used to map each interaction event in the interaction behavior sequence into a low-dimensional dense vector as an embedded representation of the interaction event, wherein the interaction behavior sequence [(object 1, operation 1), (object 2, operation 2), ..., (object n, operation n)] contains n interaction events (object, operation); the embedded representations of the n interaction events are combined into an event embedding sequence; and the semantic feature vector of the interaction behavior sequence is obtained according to the event embedding sequence.

[0059] Specifically, the data encoding unit uses the intention recognition model to encode the interaction behavior sequence in the first data set and extract the semantic feature vector of the behavior; the intention recognition model adopts a recurrent neural network (RNN) architecture based on the attention mechanism, such as a long short-term memory network (LSTM) or a gated recurrent unit network (GRU). This embodiment adopts the LSTM model, which can effectively process time series data and capture long-distance dependencies in the sequence. Given a user's interaction behavior sequence [(object 1, operation 1), (object 2, operation 2), ..., (object n, operation n)], the LSTM model first maps each interaction event (object, operation) to a low-dimensional dense vector as an embedded representation of the event. Then, the LSTM model processes the event embedding sequence in an iterative manner, updates the hidden state at each time step, and generates a hidden representation of the current event. The LSTM model also introduces an attention mechanism so that the hidden state can adaptively focus on key interaction events. Finally, the hidden state sequence output by the LSTM model is subjected to operations such as pooling to obtain the semantic feature vector of the entire interaction behavior sequence. Through unsupervised pre-training, the intent recognition model can automatically learn event embedding and semantic encoding functions from large-scale user behavior data without manually designing features. The intent recognition model makes full use of the temporal information of the behavior sequence to mine the semantic connections and dependencies between events; it adaptively focuses on important events through the attention mechanism to improve the expressiveness of semantic features. The data encoding unit learns the implicit semantic representation of user behavior through a deep learning model, laying the foundation for subsequent intent type recognition and intent evolution analysis.

[0060] For example, assume that the sequence of user interaction behaviors in the virtual mall is [(item A, browse), (item B, browse), (item C, add to cart), (item D, browse), (item C, purchase)]. The intent recognition model first maps each item and operation into an embedding vector, and then processes these vectors in sequence using the LSTM model, and focuses on key events such as (item C, add to cart) and (item C, purchase) through the attention mechanism. By pooling the hidden state of the LSTM, the semantic feature vector of the behavior sequence can be obtained, capturing the semantic information that the user is interested in item C and has the intention to purchase it.

[0061] The explicit intention recognition unit is used to calculate the similarity matrix between semantic feature vectors; construct a minimum spanning tree based on the similarity matrix; and set a distance threshold on the minimum spanning tree. , according to the distance threshold A cluster of semantic feature vectors is obtained; each cluster of semantic feature vectors is matched with a predefined intent type to obtain an explicit intent type of the cluster.

[0062] Specifically, the explicit intent recognition unit uses the intent map algorithm to cluster semantic feature vectors. The intent map algorithm can find cluster structures of any shape in the semantic feature space and adaptively determine the number of clusters and hierarchical relationships. First, the intent map algorithm calculates the similarity matrix between semantic feature vectors. Commonly used similarity metrics include cosine similarity, Euclidean distance, etc. Then, the algorithm constructs a minimum spanning tree (MST) based on the similarity matrix. By setting a suitable distance threshold on the MST, clustering hierarchies of different granularities can be obtained. Each cluster can be regarded as a semantically similar behavior cluster, reflecting an explicit interaction intention of the user. Next, the intent map algorithm matches each cluster with a predefined intent type, and uses methods such as text similarity and rule matching to find the most semantically relevant intent type. The predefined intent types can come from domain knowledge bases or expert experience, such as "product browsing" and "social interaction". Through clustering and matching, the algorithm realizes the mapping from behavioral semantics to explicit intent types. The intent map algorithm can adaptively discover behavior clusters in the semantic feature space without presetting the number of clusters; it can flexibly adjust the abstract level of intent types by obtaining multi-granular representation of intent through hierarchical clustering; it can guide intent type matching using domain knowledge base to improve the interpretability of recognition results. The explicit intent recognition unit converts the semantic features of user behavior into explicit intent types, revealing the main interaction purposes and motivations of users in the metaverse environment.

[0063] Exemplarily, for the user behavior semantic feature vectors extracted from the data encoding unit, the intent map algorithm first calculates the similarity matrix between them. Then, by setting the distance threshold on the minimum spanning tree, behavior clusters of different granularities can be obtained, such as [(browse product A, browse product D), (browse product B), (add to cart product C, purchase product C)]. Each cluster reflects an explicit intention, such as "browse products of interest", "buy decision products", etc. The algorithm further matches these clusters with predefined intent types to obtain intent type labels, such as "browse products", "purchase products".

[0064] The implicit intention recognition unit is used to establish an intention evolution model, obtain the hidden state sequence of the user interaction behavior sequence through the intention evolution model, and obtain the user's implicit intention and implicit intention type according to the hidden state sequence.

[0065] Specifically, the implicit intention recognition unit uses the intention evolution model to perform temporal modeling and reasoning on the motivations and intentions behind the user's interactive behavior. The intention evolution model is based on the hidden Markov model (HMM), which models the user's true intention as a hidden state and the observed interactive behavior as an observed state. It characterizes the dynamic evolution of the intention through the transition between states and the observation generation probability. In the HMM, there is a transition probability between hidden states, which reflects the temporal dependency between different intentions; the observation state is generated by the hidden state with a certain probability, which reflects the correspondence between the intention and the surface behavior. By estimating the model parameters of the HMM, including the initial state probability matrix, the state transition probability matrix, and the observation generation probability matrix, the intention evolution model can fit the temporal pattern of the user's interactive behavior.

[0066] In the model training phase, the forward-backward algorithm and other methods are used to estimate the model parameters by maximizing the generation probability of the observation sequence. This process can automatically learn the statistical laws of intent evolution from large-scale user behavior logs. The trained intent evolution model can be used to infer the intent of new user interaction sequences. Given an observed interaction behavior sequence, the model calculates the posterior probability distribution of the hidden state at each time step through the forward-backward algorithm, or uses the Viterbi algorithm to find the most likely hidden state sequence, thereby inferring the user's intention state at different times. By analyzing the intention state sequence, some new intent types that have not been defined in explicit intent but frequently appear, namely implicit intent, can be mined, such as "compare products" and "collect content of interest", thereby expanding the original set of intent types. In addition, the transition patterns between intent states can also be examined to find some typical intent evolution paths, such as the path from "product browsing" to "add to shopping cart" to "complete purchase", which reflects the progression of user interests and the formation process of decision-making.

[0067] Through the modeling of intent evolution, the implicit intent recognition unit can characterize user intent in the time dimension, explore the deep intent hidden under the behavioral appearance, and provide the temporal characteristics of intent changes. This helps to understand user behavior from the perspective of continuous interaction, predict users' possible future intentions, and achieve more accurate intention guidance and services. At the same time, the temporal pattern of intent reveals the general rules of users from cognition, decision-making to action in the metaverse scenario, which is instructive for optimizing interaction processes and designing intelligent services. The implicit intent recognition unit adopts an unsupervised learning paradigm, overcomes the limitations of manually defined intent, can adaptively discover new implicit intent types, and improves the flexibility and openness of intent discovery.

[0068] For example, in a virtual tourism scenario, the user's interaction log contains a series of behaviors, such as "browsing the introduction of scenic spots", "checking travel guides", "participating in virtual tours", "sharing tour experiences", etc. By training the intention evolution model, some implicit intention states can be identified from these behaviors, such as "information collection", "planning", "experience", "social sharing", etc. By analyzing the intention state sequence, it can be found that "information collection" to "planning" is a high-frequency transfer mode, revealing that users usually make careful preparations before participating in virtual tourism. The intention transfer from "experience" to "social sharing" reflects the user's desire to share the immersive tour experience with friends, reflecting the importance of virtual social interaction in tourism activities. These implicit intentions and their temporal characteristics mined from behavioral data provide an important basis for designing personalized smart guides, recommending interesting tour items, and organizing tourism social activities.

[0069] The data combination unit is used to construct a hierarchical intent type system according to the semantic relationship between intent types; map each user interaction behavior sequence to the intent type system, and assign a corresponding intent type label to each interaction event in the interaction behavior sequence; record the timestamp of each intent type occurrence to obtain the time sequence of the intent sequence; and generate a second data set according to the time sequence of the intent sequence;

[0070] Specifically, after obtaining explicit intents and explicit intent types and implicit intents and implicit intent types, these intents need to be systematically organized and integrated. First, the data combination unit constructs a hierarchical intent type system based on the semantic relationship between intent types. The system adopts a tree structure to organize intent types hierarchically from abstract to specific. For example, in the commodity purchase scenario, "commodity browsing" can be used as a parent intent, and "browsing electronic products", "browsing clothing", etc. can be used as sub-class intents. Next, each user interaction behavior sequence is mapped to the intent type system, and each interaction event in the interaction behavior sequence is assigned a corresponding intent type label. At the same time, the timestamp of each intent type instance is recorded to obtain the time sequence of the intent sequence. On this basis, each data sample in the generated second data set includes attributes such as user ID, intent type ID, intent hierarchy, and intent timestamp. In this way, a structured intent data set is obtained, which contains both the semantic information of the intent and the temporal information of the intent. By constructing a hierarchical intent system, we can achieve a systematic organization of intents of different granularities, which is convenient for overall grasp and understanding of user intents; combining intent annotation with time series information not only retains the semantic attributes of intent, but also describes the process of intent evolution over time; recording user intent in the form of structured data samples provides a unified data input format for subsequent intent mining and user analysis tasks. The data combination unit further abstracts and structures the identified user intent, so that it can be well combined with other information such as knowledge bases and user portraits, enhancing the availability and interpretability of intent analysis results. At the same time, regular intent data samples also facilitate the application of machine learning algorithms to explore deeper user behavior patterns and preference rules.

[0071] For example, for a user’s interactive behavior sequence: [(browse product A, t1), (browse product B, t2), (add to cart product C, t3), (browse product D, t4), (purchase product C, t5)]. According to the intent type system, these events can be mapped to [(browse product - electronics, t1), (browse product - electronics, t2), (add to cart, t3), (browse product - clothing, t4), (purchase product, t5)]. In the corresponding samples of the second dataset, the intent types include “browse product - electronics”, “add to cart”, “browse product - clothing”, and “purchase product”. The intent hierarchy is [“browse product”, “browse product”, “cart management”, “browse product”, and “purchase product”], and the intent time sequence is [t1, t2, t3, t4, t5]. This structured intent representation provides rich feature information for subsequent tasks such as user portrait analysis and personalized recommendation.

[0072] The second data generation unit analyzes user behavior in multiple dimensions and extracts refined user intent representations from massive interaction logs. The intent recognition model uses deep learning technology to automatically learn semantic patterns in behavioral data, overcoming the limitations of manually defined rules. On this basis, the intent map algorithm performs clustering and induction, refines the identified intent, and improves the coverage and accuracy of the intent system. The introduction of the intent evolution model further explores the evolution of motivations behind user behavior and discovers implicit intentions that are difficult to directly observe but have an important impact on user decisions. Through the hierarchical organization and temporal representation of intent, the second data generation unit constructs a comprehensive and three-dimensional user intent portrait, which not only reflects the user's immediate demands in the metaverse, but also reveals the long-term evolution of user interests and preferences, which is of great value for applications such as personalized recommendations and intelligent interactions.

[0073] The interaction intention knowledge graph construction module is used to construct a user-centric interaction intention knowledge graph based on the second data set;

[0074] The interactive intention knowledge graph construction module includes: a graph mapping unit, a node association unit and an attribute attachment unit;

[0075] The graph mapping unit is used to map the intent type, intent hierarchy, and intent sequence in the second data set into nodes and edges in the knowledge graph;

[0076] In the graph mapping unit, the intent types, intent levels, and intent sequences in the second data set are mapped to nodes and edges in the knowledge graph, including: mapping intent types to intent type nodes in the knowledge graph; mapping intent instances at different time points in the intent sequence in the second data set to intent instance nodes under corresponding type nodes; constructing intent sequence edges between intent instance nodes based on intent sequence; mapping each user to a user node in the knowledge graph; establishing subclass relationship edges between corresponding intent type nodes based on the hierarchical relationship between intent types; and establishing instance relationship edges between intent instance nodes and corresponding intent type nodes.

[0077] Specifically, in the graph mapping unit, the intent information extracted from the second data set is converted into the basic building unit of the knowledge graph. The intent type is introduced into the graph as the top-level node to form a hierarchical organization of the intent concept. Under each intent type, corresponding sub-level intent nodes are created according to the specific intent instances at different time points in the intent sequence. These intent nodes are connected by directed edges, and the direction of the edge represents the temporal sequence of the intent, thereby encoding the dynamic evolution characteristics of the intent in the graph. By introducing the association between user nodes and intent nodes, a mapping between individual users and group intentions is established. Through subclass relationship edges and instance relationship edges, the knowledge graph embodies the semantic associations between intents of different granularities in the form of an ontology.

[0078] The core of the graph mapping unit is to achieve the conversion from unstructured intent data to structured knowledge representation. Integrate the scattered intent analysis results into a semantically coherent graph to make it more logical and interpretable. The graph stores intent knowledge in the form of a graph database, supporting flexible association queries and reasoning analysis. The graph-based intent representation overcomes the semantic opacity of vector representation. The intent nodes are displayed intuitively and the intent relationships are clear at a glance. Using the graph as an intermediary, it is also convenient to interconnect intent knowledge with other domain knowledge bases to expand the multi-semantics of intent.

[0079] For example, in a virtual shopping scenario, a shopping intention knowledge graph can be constructed through the graph mapping unit, and "product browsing", "add to shopping cart", "product purchase" and other top-level intention type nodes are formed; a user's shopping behavior sequence is analyzed and extracted, and the key intention instance nodes are identified, such as "browse product A", "browse product B", "add to cart product C", "purchase product C", etc., and intention timing edges with arrows are established based on their temporal dependencies. The user node is connected to these intention instance nodes, indicating that this is a specific shopping intention path for the user. Different types of product nodes can also be associated with the intention instance nodes corresponding to the browsing and purchasing behaviors. In this way, through the intention knowledge graph, the intention context behind the user behavior can be interpreted from a structured semantic perspective, thereby optimizing product design and service strategies.

[0080] By embedding the intent time series into the graph structure, the dynamic characteristics of user intent can be fully characterized and measured. On the one hand, the evolution trajectory of a single user's intent under multiple interactions can be characterized through the intent node sequence. On the other hand, the similarity of different user intent paths can also be calculated based on the topological structure of the graph, such as using a random walk algorithm to measure the relative distance of different paths in the graph. This provides a new perspective for understanding the overall intent preferences and evolution patterns of group users. The construction of the intent knowledge graph expands user behavior analysis from static feature extraction at a single time point to dynamic association representation across time domains, allowing behavioral data to better serve business analysis and decision-making.

[0081] The node association unit is used to associate a series of intention instance nodes of the same user at different time points with the user node of the user, forming a star topology subgraph with the user node as the center and surrounded by the intention instance nodes.

[0082] The attribute adding unit is used to add attribute information to the nodes and edges in the knowledge graph to obtain the final interaction intention knowledge graph;

[0083] In the attribute adding unit, attribute information is added to the nodes and edges in the knowledge graph to obtain the final interactive intention knowledge graph, including: for the intent type node, adding statistical attributes including word frequency and TF-IDF; for the intent instance node, adding contextual attributes including timestamp, geographic location, and emotional tendency; for the intent time sequence edge, adding dynamic attributes including transition probability and time decay coefficient; for the user node, associating the user's demographic attributes and preference attributes;

[0084] Specifically, the attribute attachment unit further enriches the semantic dimension of the interactive intent knowledge graph, and attaches a variety of attribute information to the nodes and edges in the graph. These attributes characterize the characteristics of the intent elements from different aspects. For example, statistical attributes reflect the popularity distribution of intent topics, contextual attributes reflect the external environment in which the intent occurs, and dynamic attributes quantify the temporal dependency strength of the intent. Combining these attributes with the structural information of the graph, the representation of intent knowledge is more three-dimensional and comprehensive.

[0085] Taking the intent type node as an example, term frequency refers to the frequency of a certain word appearing in all user behavior records mapped by the intent type. Intuitively, it reflects the degree of user attention to the object or concept corresponding to the word in the relevant intent scenario. TF-IDF (Term Frequency-Inverse Document Frequency) further considers the distribution of words in the entire corpus on the basis of term frequency. If a word appears frequently under one intent type but rarely in other intents, its TF-IDF value will be very high, indicating that it is a significant feature word of the intent. For example, the TF-IDF value of the word "duty-free" in the "overseas shopping" intent may be much higher than that in the "local shopping" intent. The statistical properties of the intent node help to summarize the significant objects and topics that the intent focuses on, making it easier to quickly locate and interpret the core connotation of the intent.

[0086] The contextual attributes of intent instance nodes, such as geographic location, can reveal the spatial distribution characteristics of user intent. Users in different regions may show significant preference differences for the same type of intent. Taking the "scenic spot tour" intent as an example, users in Beijing may pay more attention to cultural relics such as the Forbidden City and the Great Wall, while users in Sanya may prefer leisure experiences such as beaches and diving. Capturing this regional characteristic is of great significance for carrying out precise operations and localized services. Emotional tendency describes the psychological state of users under specific intents from the perspective of subjective feelings. It is common for some comment interaction intentions to express positive emotions, while consultation and complaint intentions are often accompanied by negative emotions. Through the correlation analysis of intent and emotion, we can gain insight into the emotional connection between users and products and optimize the interaction design of key nodes.

[0087] The transition probability on the time-series edge of the intent describes the probability of following the target intent node immediately after the source intent node. This indicator learns the statistical laws of intent migration from a large sample of user behaviors and quantifies the strength of the dependency between two adjacent intent instances. The time decay coefficient further considers the relationship between intent migration and time distance. Intuitively, two intents with a short time interval will have a higher correlation than two intent pairs with a long time interval. Taking the game scenario as an example, if a user accesses the equipment enhancement interface immediately after a dungeon battle, the transition probability is usually higher than if the user enhances the equipment long after the battle ends. Combining the transition probability and the time decay coefficient, a quantitative indicator system for intent association can be established to accurately characterize the dependency effect between complex intents from a time-series perspective.

[0088] The attribute information of user nodes, such as demographic variables and preference labels, explains their intended behavior from the user's own portrait characteristics. Comparative analysis of the intention patterns of different groups of people in the knowledge graph can reveal valuable behavioral differences and commonalities. For example, by filtering by age attributes, it may be found that young groups are more active in social entertainment intentions, while middle-aged and elderly groups may be more concerned with health consultation intentions. Combining these conclusions with the user portrait label system can achieve refined identification of user intentions and personalized contact with thousands of people.

[0089] In summary, the attribute addition unit strengthens the expressiveness of intent knowledge representation by enriching the semantic attributes of the knowledge graph. The organic integration of multi-dimensional attributes and graph structure enables the knowledge graph to describe intent behavior from a full range of perspectives such as statistics, time series, emotions, and user characteristics, providing a powerful analysis tool for user insights and business decisions. The association analysis from intent to user also provides important intent features for personalized recommendations and services. Graph-based representation learning and mining reasoning technology will further unleash the value of intent knowledge and enable intent recognition and prediction with intelligent algorithms.

[0090] For example, in the social media scenario, a user intention knowledge graph is constructed, in which "browsing friends' dynamics", "posting status", "interactive comments", "private message chat" and other behaviors form intention type nodes. After analysis, user behaviors such as "Zhang San browses friend Li Si's daily dynamics", "Zhang San comments on Li Si's photos", and "Zhang San sends Li Si a private message" form specific intention instance nodes and are connected according to the temporal relationship. Statistical attributes such as word frequency and TF-IDF reveal Zhang San's focus on friends' dynamics, such as the prominence of high-frequency words such as "travel" and "food". Contextual attributes such as the temporal and geographical distribution of interactive behaviors reflect Zhang San's social active time period and area. Transition probability and time decay coefficient characterize the role of different interactive intentions, such as Zhang San tends to interact again soon after commenting. Portrait labels such as age, gender, hobbies and interests on user nodes also provide an individual perspective for interpreting intentions. The convergence analysis of social intentions of multiple users can reveal the interaction patterns and popular topics of the group. The application of these intention knowledge will eventually serve the construction of intelligent social platforms and realize information push and interactive guidance for thousands of people.

[0091] The interactive intention knowledge graph construction module uses knowledge graph construction technology to semantically integrate user intention data from different sources and in different forms to form a unified, machine-readable knowledge representation. In the graph construction process, the structured intent information in the second data set needs to be extracted as entities and relationships. Among them, different levels of intent types are modeled as hierarchical type nodes, such as "entertainment intent" contains sub-class nodes such as "game intent" and "movie intent"; specific intents identified at different time points are modeled as instance nodes, such as "game intent at time t1" and "game intent at time t2". The relationship between nodes reflects the semantic connection between intents, such as the upper and lower relationships of "include" and "belong" between intent types, and the temporal relationship or association relationship of "adjacent to each other" and "co-occurrence" between intent instances. A series of intent nodes of the same user at different times are connected to the user node to form a star topology structure with the user node as the center and the intent nodes surrounding it, which intuitively presents the personalized intention distribution of different users. The constructed intent knowledge graph is semantically enriched.

[0092] Through the interactive intention knowledge graph construction module, the fragmented intentions of users in the metaverse scenario are connected into a semantically coherent knowledge graph, revealing the multi-level semantic structure and dynamic evolution law of user intentions. Mapping heterogeneous intention data to a unified graph space makes it more understandable and computable, and facilitates automatic application to intelligent services. The graph-based intention representation has explicit semantics and structure, and computers can perform complex cognitive tasks such as semantic understanding and relational reasoning through graph algorithms. Organizing intention information with user nodes as anchors directly establishes the association between users and intentions, providing a basis for personalized, context-aware intelligent interaction. The knowledge graph also provides flexible semantic query and data integration mechanisms, which can efficiently retrieve and aggregate background knowledge related to the current intention from massive heterogeneous data. This knowledge-driven intelligent interaction method allows the metaverse platform to understand the ins and outs of user intentions like humans and infer the real needs of users in a specific context. For example, when a user browses a movie, the platform can not only identify the "movie-watching intention", but also infer based on the relationship in the graph that the intention may be triggered by the previous "browse movie reviews" intention, and then actively recommend relevant movie review articles to the user. When the user finishes watching the movie, the platform can also predict the user's social intention to "share the movie-watching experience" based on the timing of the intention, and prompt the user to perform subsequent actions such as rating and writing movie reviews. The knowledge graph allows machines to understand people's deep intentions through superficial interactive behaviors. By embedding intentions into broader semantic associations for understanding, it not only improves the accuracy of intent recognition, but also makes intent responses more in line with people's behavioral habits and ways of thinking.

[0093] For example, assume that user A has the following interactive behavior sequence in the virtual mall: browse the mobile phone product page (t1), browse the mobile phone details (t2), add to the shopping cart (t3), browse the VR helmet product page (t4), and place an order to buy a VR helmet (t5). Through intent recognition, the user's intention type at each moment can be obtained, such as "browse products" (t1, t2, t4), "add to shopping cart" (t3), and "buy products" (t5). These intention data are integrated into the knowledge graph, and the intention nodes of the same type are automatically aggregated. For example, the browsing behavior of different products corresponds to the same "browse products" node. A series of intention nodes of a single user are associated with the user node to form a personalized intention subgraph. Based on the intention time sequence path from "browse products" to "add to shopping cart" and then to "buy products" in the graph, the platform can predict the user's purchase intention and push marketing information such as product coupons at time t2. In addition, based on the associated edges of different product intention nodes, the platform finds that "browse mobile phones" and "browse VR helmets" have a co-occurrence relationship, and can actively recommend mobile phone products that have been viewed before when users browse VR helmets. These personalized services based on intent knowledge do not require manual configuration of rules. Instead, they automatically mine behavioral patterns from the intent graphs of massive users, giving the platform the intelligence to proactively analyze and predict user needs.

[0094] In summary, the interactive intention knowledge graph construction module uses the knowledge graph to perform semantic abstraction and association mining on the intention data, and builds a cognitive bridge connecting users, intentions, and interactive behaviors. The graph-centric intention calculation method allows the machine to understand the real needs of users in the scene, and guide users to complete the interactive goals in an intelligent and natural way, allowing the Metaverse platform to break away from the rigid interactive process and enter the intelligent era driven by user intentions.

[0095] The vertical domain knowledge graph construction module is used to identify key vertical domain knowledge for different metaverse application scenarios and construct a vertical domain knowledge graph based on the vertical domain knowledge;

[0096] The vertical field knowledge graph construction module includes a knowledge identification unit and a graph construction unit;

[0097] The knowledge identification unit is used to determine the key vertical fields of the metaverse application scenarios; for each vertical field, it collects structured knowledge including domain ontology, knowledge base and industry dictionary and unstructured knowledge including domain documents and user comments; and constructs vertical domain knowledge from structured knowledge and unstructured knowledge.

[0098] Specifically, the knowledge identification unit identifies vertical domain knowledge that is crucial to intelligent services by analyzing the characteristics of different metaverse application scenarios. Taking virtual exhibitions as an example, exhibit information, exhibitor introductions, industry dynamics, etc. are key elements that constitute the exhibition knowledge graph. In the game scenario, game mechanisms such as game rules, character attributes, and equipment effects are indispensable professional knowledge. For each vertical field, the knowledge identification unit makes full use of high-quality structured knowledge resources such as domain ontology and industry dictionary to quickly build the basic skeleton of domain knowledge. At the same time, the knowledge identification unit also performs semantic analysis and knowledge extraction on domain-related unstructured data. Taking user comments as an example, it can identify implicit knowledge such as users' concerns and pain points about virtual exhibits, and supplement item attributes. By comprehensively utilizing structured and unstructured data, the knowledge identification unit realizes the comprehensive collection and refinement of vertical domain knowledge, providing a solid knowledge foundation for knowledge-based intelligent services. Making full use of vertical domain knowledge helps to enhance the professionalism and pertinence of intelligent services, and provide personalized services that meet industry needs and user cognition.

[0099] The graph construction unit is used to define the core concepts, entities, entity types, and attributes of the vertical field, map the core concepts, entities, entity types, and attributes of the vertical field into nodes in the vertical field knowledge graph; map the relationships between entities in the vertical field into edges in the knowledge graph; establish subclass relationship edges between corresponding entity type nodes based on the hierarchical classification system of the vertical field; establish logical rule edges between related entity nodes; map events within the vertical field into event nodes in the vertical field knowledge graph, and establish association edges with related entity nodes.

[0100] For example, in the field of e-commerce, products, brands, categories, etc. can be used as entity nodes; relationships such as "product-brand", "product-category", "product-attribute" can be used as edges, and relationships such as "user-product" and "user-comment" can also be used as edges; a subclass relationship edge is established between "electronic products" and "mobile phones" to indicate that the latter is a subclass of the former; rule edges are used to represent e-commerce common sense such as "a certain brand belongs to a certain category" and "two certain products are often purchased together"; each order can be used as an event node, whose attributes include user characteristics, purchased products, payment amount, logistics status, etc., and it is connected to related product, user and other entity nodes.

[0101] Through the graph construction unit, the structured, semi-structured and unstructured knowledge in vertical fields can be integrated into a unified graph representation. This graph-based knowledge representation can fully explore the intrinsic connections between concepts and entities and reveal the hierarchical structure and logical rules of domain knowledge. At the same time, entities, relationships, events, etc. in the graph can be associated with more external information, such as text descriptions, multimedia content, etc., thereby further enriching the semantics of knowledge representation.

[0102] Building a vertical domain knowledge graph provides a unified and standardized form of knowledge organization, which is conducive to the management, retrieval and application of knowledge. Traditional relational databases are difficult to handle complex semantic relationships between entities, while knowledge graphs model entities and their relationships in the form of graphs, which is more in line with human cognitive habits. Knowledge graphs can mine semantic patterns and association rules hidden in entities and relationships through graph algorithms and graph embedding technologies. For example, through graph representation learning, semantic similarities between concepts can be automatically discovered, and reasoning such as concept clustering, generalization, and refinement can be achieved. Through graph path analysis, multi-hop associations between entities can be discovered, and association-based recommendations and predictions can be achieved, which helps to discover new knowledge and reveal deep domain rules. Through the construction of knowledge graphs, implicit knowledge such as the experience and insights of domain experts can be made explicit, which is conducive to the inheritance and sharing of knowledge. With the continuous improvement and updating of knowledge graphs, it can become an authoritative reference for knowledge in the field, providing explainable and reliable knowledge support for various intelligent applications.

[0103] For example, in the field of e-commerce, a large amount of transaction data is scattered in various business systems, lacking a global perspective. By constructing an e-commerce knowledge graph, key concepts such as goods, brands, categories, and users can be extracted to form structured entity nodes; and based on product attributes, user behavior, transaction records, etc., the associations, purchase rules, preference patterns, etc. between these concepts can be mined to form rich relationship edges. The e-commerce knowledge graph can be used to support e-commerce scenarios such as personalized recommendations, intelligent search, and customer portraits. For example, when a user enters a keyword in the search box, the platform can perform semantic matching on the e-commerce knowledge graph, understand the user's intent, and return the most relevant products or categories. For another example, analyze the user's historical browsing and purchasing behavior in the knowledge graph, discover their preferences, and recommend other products that similar users like to improve transaction conversion rates.

[0104] In short, vertical domain knowledge graphs make the representation and application of knowledge more intelligent by mapping, semanticizing, and associating domain knowledge, which is conducive to improving the professional level and service quality of metaverse applications; the integration with user intention graphs further expands the application scenarios and implementation paths of the metaverse in vertical industries; this "AI+knowledge" integration method represents an important direction for the future development of the metaverse.

[0105] The knowledge graph fusion module is used to fuse the vertical domain knowledge graph with the interactive intention knowledge graph to obtain a fused heterogeneous knowledge graph;

[0106] The knowledge graph fusion module includes an embedding vector generation unit, an embedding vector optimization unit and a graph fusion unit;

[0107] The embedding vector generation unit is used to adopt a graph neural network to learn the dense embedding representation of nodes and edges in a vertical domain knowledge graph, and generate an initial graph embedding vector including a node embedding vector and an edge embedding vector; the node embedding vector includes an entity node embedding vector.

[0108] The embedding vector optimization unit is used to propagate the node embedding vectors in the initial graph embedding vector through the graph convolutional network, so that entities with similar semantics or structures in the graph obtain similar node embedding vectors, and use the similar node embedding vectors to update the node embedding vectors in the initial graph embedding vector to obtain an optimized graph embedding vector.

[0109] For example, assume that a knowledge graph in the e-commerce field is constructed in the vertical field knowledge graph construction module. The graph contains different types of nodes such as products, brands, categories, and relationship edges between them, such as "product-belongs-category", "product-belongs-brand", etc.; in the embedding vector generation unit, the initial graph embedding vectors of all nodes and edges are obtained through the graph neural network.

[0110] For example:

[0111] Node 1: “Mobile phone” category node, the initial graph embedding vector is [0.2,-0.1,0.5,];

[0112] Node 2: “Apple” brand node, the initial graph embedding vector is [0.3, 0.2, 0.1,];

[0113] Node 3: “iPhone12” product node, the initial graph embedding vector is [-0.1, 0.4, -0.3,];

[0114] Node 4: “Huawei” brand node, the initial graph embedding vector is [0.6,-0.5,0.2];

[0115] Node 5: “Huawei P40” product node, the initial graph embedding vector is [0.1,-0.2,0.4,];

[0116] Edge 1: The “belongs to” relationship edge from node 3 to node 1, with the initial graph embedding vector being [0.2, 0.3, -0.1,];

[0117] Edge 2: The “belongs to” relationship edge from node 3 to node 2, with the initial graph embedding vector being [-0.1, 0.5, 0.2,];

[0118] Edge 3: The “belongs to” relationship edge from node 5 to node 1, with the initial graph embedding vector being [0.3, 0.1, 0.4,];

[0119] Edge 4: The “belongs to” relationship edge from node 5 to node 4, with the initial graph embedding vector being [0.5,-0.1,-0.2];

[0120] Next, in the embedding vector optimization unit, the graph convolutional network (GCN) updates the node embedding according to the edge connection relationship between nodes; taking node 3 as an example:

[0121] Find the neighbor nodes of node 3, including node 1 (connected by edge 1) and node 2 (connected by edge 2);

[0122] The node embedding vectors of node 1 and node 2 are transformed according to the corresponding edge embeddings, and then weighted summed to obtain the aggregated representation of the neighbor information;

[0123] Perform nonlinear transformation on the embedding of node 3 itself and the neighbor aggregation representation to obtain the updated embedding vector of node 3;

[0124] Similarly, the graph convolutional network also updates the embeddings of other nodes. After multiple iterations, the updated node embeddings will propagate the graph structure information, so that the related nodes obtain similar embedding representations.

[0125] For example, iPhone 12 and Huawei P40 belong to the same “mobile phone” category, and after GCN embedding propagation, their node embeddings will be closer. There is also a “product-brand” relationship edge between iPhone 12 and the Apple brand, so their node embeddings will also be similar.

[0126] The embedding vector optimization unit propagates and updates the node embedding vectors in the initial graph embedding vector through the graph convolutional network, so that entities with similar semantics or structures in the graph obtain similar node embedding vectors. This process significantly enhances the ability of semantic expression, making similar nodes closer in the embedding space, thereby better capturing the semantic relationship between entities. By integrating the information of adjacent nodes, the node embedding vector not only contains its own characteristics, but also incorporates the attributes of surrounding nodes, improving the generalization ability of the model in processing complex relationships. This optimization helps to improve prediction accuracy, enabling the model to more accurately reflect the potential connections between entities, thereby improving the performance of downstream tasks such as classification and recommendation. In addition, by realizing the fusion of knowledge from different fields in the graph embedding space, implicit semantic connections across fields can be naturally discovered to support more complex intelligent services. By reducing data sparsity, the optimized node embedding makes the processing of large-scale graph data more efficient, laying a solid foundation for intelligent services in complex application scenarios.

[0127] The graph fusion unit is used to associate the optimized graph embedding vector of the vertical field with the nodes in the interactive intent knowledge graph to obtain a fused heterogeneous knowledge graph; the nodes in the interactive intent knowledge graph include user nodes and intent instance nodes.

[0128] In the graph fusion unit, associating the optimized graph embedding vector of the vertical field with the node in the interactive intention knowledge graph includes: obtaining the node embedding vector of the interactive intention knowledge graph; calculating the similarity between the node embedding vector of the interactive intention knowledge graph and the node embedding vector of the vertical field knowledge graph, and calculating the similarity of the node embedding vector greater than the similarity threshold. The nodes of the interactive intention knowledge graph and the nodes of the vertical domain knowledge graph establish semantic associations between cross-graph nodes; the semantic associations between cross-graph nodes are used to construct a node alignment loss function; according to the node alignment loss function and the structure preservation loss of a single graph, a comprehensive optimization function is established, and the fused heterogeneous knowledge graph is obtained by minimizing the comprehensive loss function; the structure preservation loss of the single graph includes the structure preservation loss function of the interactive intention knowledge graph and the structure preservation loss function of the vertical domain knowledge graph.

[0129] In the graph fusion unit, a method for calculating the similarity between the node embedding vector of the interaction intention knowledge graph and the node embedding vector of the vertical domain knowledge graph includes:

[0130]

[0131] in:

[0132] : Nodes in the interactive intention knowledge graph Nodes in the vertical domain knowledge graph similarity;

[0133] : Nodes in the interactive intention knowledge graph The embedding vector of

[0134] :Nodes in the vertical domain knowledge graph The embedding vector of

[0135] :node The embedding vector Dimensional component;

[0136] :node The embedding vector Dimensional component;

[0137] : The dimension of the embedding vector;

[0138] : No. The weight coefficient of the dimension feature;

[0139] : No. Dimensional feature scale parameter;

[0140] : An exponential function with the natural constant e as the base;

[0141] and It is the node embedding representation learned by applying graph embedding algorithms (such as GCN, GAT, etc.) to their respective knowledge graphs; weight coefficient and scale parameter It can be tuned by cross-validation or grid search, or it can be set as a priori. The formula calculates the weighted exponential similarity of the two node embedding vectors in each dimension, and then normalizes the similarity of all dimensions to obtain the final node similarity. Exponential similarity function The two nodes are measured in The similarity of the two dimensional features decreases exponentially as the feature difference increases. Used to adjust the importance of features of different dimensions, allowing flexibility in emphasizing or ignoring certain features. Controls the rate of exponential decay, The larger the value, the slower the similarity decays and the higher the tolerance to feature differences. The denominator performs a Normalization eliminates the influence of the embedding vector length and makes the similarity value more stable.

[0142] The value range is The larger the value, the more nodes and The higher the semantic similarity of . When the embedding values ​​of two nodes in each dimension are close, that is, Smaller, The value of will be larger. When two nodes are in an important dimension (i.e. When the embedding values ​​on the larger dimension are close, The value of will be relatively larger. The larger it is, the less sensitive the similarity is to feature differences. The value of is relatively larger; on the contrary, The smaller it is, the more sensitive the similarity is to feature differences. The value of is relatively smaller. The normalization factor makes The size of is not affected by the length of the embedding vector and only depends on the proximity in the vector direction.

[0143] The formula introduces weight and scale parameters. By adjusting these parameters, important features can be emphasized, minor features can be ignored, and the sensitivity of similarity to feature differences can be controlled. At the same time, the exponential similarity function imposes a stronger penalty on features with larger differences, making the similarity more discriminative. This adaptive similarity calculation method helps to more accurately characterize the semantic associations between nodes and provides a more reliable basis for the alignment and fusion of knowledge graphs.

[0144] In the graph fusion unit, the node alignment loss function is:

[0145]

[0146] in:

[0147] : is the set of node pairs to be aligned;

[0148] : is the cross-graph node pair to be aligned, i'' represents the node in the interaction intention knowledge graph, and j'' represents the node in the vertical domain knowledge graph;

[0149] : is the sigmoid function;

[0150] : are unaligned cross-graph node pairs, i.e., negative samples. These nodes are semantically irrelevant. Represents a node in the interaction intention knowledge graph, Represents a node in a vertical domain knowledge graph;

[0151] : is the number of negative samples, which controls the number of negative samples;

[0152] : is a negative sampling distribution, used to sample unaligned node pairs ;

[0153] :Indicates that according to the distribution Sampling is performed, and the negative sample pairs obtained are ;

[0154] : Indicates the distribution of negative samples All negative sample pairs sampled in The expected value of this value measures the similarity of unaligned node pairs. Average performance on

[0155] and It is a hyperparameter that can be adjusted according to the characteristics of the dataset and the experimental results. The first Maximize the similarity of the node pairs to be aligned so that they are as close as possible in the embedding space; the second term Minimize the similarity of unaligned node pairs so that they are as far apart as possible in the embedding space; by minimizing the entire loss function, semantic alignment of nodes across graphs can be achieved, mapping semantically related nodes to similar areas in the embedding space. Similarity When it is higher, the value of the first term is smaller and the loss function value is smaller; when the misaligned nodes are Similarity When it is lower, the value of the second term is smaller and the loss function value is smaller; The larger the value, the higher the weight of negative samples and the greater the penalty for misaligned node pairs.

[0156] The node alignment loss function maximizes the similarity of the node pairs to be aligned and minimizes the similarity of the unaligned node pairs, so as to bring semantically related nodes closer to each other in the embedding space. By aligning nodes with different intents, semantically contradictory node mappings are reduced, thereby helping to maintain the semantic consistency of intent representation, which ensures that the nodes in the knowledge graph remain coherent at multiple levels of semantics; the alignment loss function minimizes the deviation between different nodes, ensuring that nodes in different fields and application scenarios can be seamlessly integrated, reducing information loss, and the platform can better handle the changes in intent in different users and scenarios, and improve generalization capabilities; in multi-domain applications, the node alignment loss function can help the intent nodes in different fields to be effectively connected and integrated, and the knowledge graph can be dynamically expanded and updated to support the growing user intent and behavior data, and support cross-domain intelligent services; by ensuring node alignment, the platform can more accurately identify the user's explicit and implicit intent, thereby improving the accuracy of recommendations and interactions.

[0157] In the graph fusion unit, the comprehensive optimization function is:

[0158]

[0159] : Node alignment loss function, used to achieve semantic alignment of nodes across graphs;

[0160] : Structure-preserving loss function for interactive intention knowledge graph;

[0161] : Structure-preserving loss function for vertical domain knowledge graph;

[0162] : Balance factor 1, used to control the weight of the structure preservation loss of the interaction intention knowledge graph;

[0163] : Balancing factor two, used to control the weight of the structure-preserving loss of the vertical domain knowledge graph.

[0164] Structure-preserving loss function for interactive intention knowledge graph And the structure-preserving loss function of the vertical domain knowledge graph The distance loss of the triplet loss function TransE, which is commonly used in knowledge graph embedding, is adopted. The distance loss of TransE is common knowledge in the art and will not be described in detail in this embodiment.

[0165] and It is a hyperparameter that can be tuned by methods such as cross-validation or grid search; and The value of affects the weights of different loss items, thus affecting the quality of the final fusion map. and When is larger, the importance of maintaining the structure of a single graph increases, and the fused graph is more inclined to retain the structural features of the original graph; and When it is smaller, the importance of cross-graph node alignment increases, and the fusion graph is more inclined to achieve semantic fusion; by balancing and The value of can achieve a compromise between cross-graph semantic alignment and single-graph structure preservation, and obtain a fused graph that takes both aspects into consideration.

[0166] By minimizing the comprehensive loss function, we can simultaneously achieve semantic alignment of cross-graph nodes and maintain the structure of a single graph, and obtain a fused high-quality heterogeneous knowledge graph; the node alignment loss function ensures that semantically related nodes are close to each other in the embedding space, achieving semantic fusion across graphs; the graph structure preservation loss function ensures that the inherent structure of a single graph is preserved in the embedding space, maintaining the semantic consistency of the graph.

[0167] In the joint optimization process, the node alignment loss across graphs and the structure preservation loss of a single graph constrain each other and jointly shape the form of the embedding space. Through iterative optimization, the node embeddings of the two graphs are "attracted" to the common semantic subspace to form a coherent knowledge representation. After the optimization is completed, the aligned node embeddings and cross-graph edge sets are integrated into a unified heterogeneous knowledge graph, which completes the fusion of the graphs. In the fused heterogeneous knowledge graph, knowledge from different fields is organically combined, expanding the breadth and depth of knowledge representation. With the intention node as the hub, user portraits are seamlessly connected with vertical field features to build a cross-domain knowledge association network. This fusion method breaks through the domain boundaries of knowledge and realizes the deep interaction between user intent and scene semantics. The consistent and rich semantic representation formed by the fusion of knowledge graphs will provide more powerful knowledge support for cross-domain intelligent decision-making and services.

[0168] The intelligent interaction module is used to build a vertical AI model based on the fused heterogeneous knowledge graph, and apply the vertical AI model to understand and predict user intentions to achieve intelligent personalized services;

[0169] The intelligent interaction module includes a vertical AI model building unit and an intelligent decision-making unit;

[0170] The vertical AI model construction unit is used to extract user intention features and domain knowledge features based on the fused heterogeneous knowledge graph, combine the user intention features and the domain knowledge features, and construct a training data set for the vertical AI model; construct a vertical AI model, and use the training data set to perform end-to-end training and optimization on the vertical AI model; the vertical AI model is a deep learning model with a graph neural network encoder and an attention mechanism decoder structure.

[0171] In the vertical AI model construction unit, the end-to-end training and optimization of the vertical AI model using the training data set includes: using the user intention features in the training data set to input the graph neural network encoder to obtain the encoded user intention representation; using the domain knowledge features in the training data set to input the graph neural network encoder to obtain the encoded domain knowledge representation; applying the attention mechanism on the encoded user intention representation and domain knowledge representation, calculating the intention-knowledge interaction weight, and obtaining the intention-knowledge combination representation; inputting the intention-knowledge combination representation into the decoder, decoding to generate personalized service decisions, and calculating the loss function of the decision; using the stochastic gradient descent algorithm to minimize the loss function of the decision, and performing end-to-end training and optimization on the parameters of the vertical AI model.

[0172] Specifically, the vertical AI model construction unit builds an AI model for a specific vertical field based on the fused heterogeneous knowledge graph to achieve user intent understanding and personalized services. User intent features include historical intent sequences and current intents; domain knowledge features include domain concepts, entities, and semantic associations related to the current scenario; first, training data is extracted from the heterogeneous knowledge graph, and the historical intent sequence of each user is obtained based on the intent knowledge graph to characterize the user's interactive behavior pattern in the metaverse; at the same time, the intent node corresponding to the user's current interactive event is extracted as the target of intent understanding. Based on the vertical domain knowledge graph, according to the current interactive scenario, the relevant domain concepts, entities and their semantic associations are identified as background knowledge constraints. The user intent features and domain knowledge features are combined in a certain way to form training samples and construct a training set. The sample can be in the form of a fixed-length feature vector or graph structured data to meet different model input requirements. The constructed training set integrates the temporal pattern of user interaction behavior on the one hand, and contains scene-related structured knowledge on the other hand, which can simultaneously train the model's intent understanding ability and domain specialization ability, thereby improving the accuracy of personalized services. This data construction method makes good use of the semantic correlation information of heterogeneous knowledge graphs and provides rich feature support for vertical field AI models.

[0173] Considering that the fused knowledge graph has a complex heterogeneous structure, traditional vector-based or sequence-based deep learning models are difficult to effectively model the semantic dependencies therein. Therefore, the vertical AI model construction unit adopts an encoding and decoding framework based on graph neural network (GNN) and attention mechanism (Attention) as the basic architecture of the vertical AI model. Among them, the encoder part uses graph neural network to represent heterogeneous knowledge graphs. Graph neural network updates the feature representation of nodes by recursively aggregating the neighbor information of nodes, and can adaptively fuse the relational structure in the graph. Common graph neural network models include graph convolutional network (GCN), graph attention network (GAT), etc. The encoder maps the knowledge graph into a low-dimensional dense node embedding vector, which not only retains the original attribute information of the node, but also incorporates the association pattern of the graph structure.

[0174] The decoder part uses the attention mechanism to combine the encoded user intent nodes and domain knowledge nodes to achieve intent understanding and service generation. The attention mechanism enables the decoder to selectively focus on the nodes most relevant to the current intent or decision by calculating the correlation weights between different node embeddings. This soft information fusion method enhances the expressiveness of the model, enabling it to flexibly adapt to the interactive relationship between intent and knowledge in different scenarios. The two-way attention calculation, on the one hand, selects the concepts and instances most relevant to the current intent from the knowledge graph to provide explanations and support; on the other hand, it uses the temporal association of intent to guide subsequent service decisions and achieve context-related personalization. The combination of the GNN encoder and the attention decoder can not only make full use of the heterogeneous information of the fused knowledge graph, but also flexibly model the interaction mode of intent-knowledge, thereby enhancing the generalization and expressiveness of the vertical AI model. This provides a powerful model foundation for intelligent services in the complex and changing metaverse environment.

[0175] The constructed training dataset is used to conduct end-to-end supervised learning of the vertical AI model. The training samples are mapped to the intent embedding space and knowledge embedding space through the GNN encoder. The encoding process makes full use of the topological structure and node attributes of the graph data to learn structured intent representation and knowledge representation. Based on the encoding, the attention mechanism obtains the combined weight of soft matching by calculating the correlation between the intent nodes and the knowledge nodes. The calculation of the weight comprehensively considers the similarity of the node semantics and the tightness of the connection structure. The combined intent-knowledge representation not only retains the original feature information, but also emphasizes the impact of cross-domain interaction. The combined representation is used as the input of the decoder to generate a personalized service decision sequence, which is matched with the target service sequence in the sample, and the generation loss is calculated to guide model optimization. Optimization algorithms such as stochastic gradient descent are used to adjust the parameters of the GNN and attention mechanism through back propagation to minimize the decision loss. This process utilizes the end-to-end data-driven paradigm, enabling the model to automatically learn the association rules between intent understanding and service decision-making from large-scale heterogeneous knowledge graphs without manually designing features. Through continuous iteration of data sets and optimization of model parameters, vertical AI models can gradually improve their ability to understand complex intents and the level of decision-making for personalized services, and adapt to the diversity and dynamics of the metaverse environment. At the same time, the intent-knowledge interaction patterns learned by the model can be migrated and reused across scenarios, reflecting the role of the introduction of vertical domain knowledge in enhancing generalization capabilities.

[0176] The intelligent decision-making unit is used to apply the vertical AI model after training and optimization to realize intelligent personalized services for different vertical scenarios of the Metaverse and enhance the user interaction experience;

[0177] Specifically, the intelligent decision-making unit obtains the user's real-time interactive behavior data in the Metaverse, extracts the current intention features, and inputs them into the vertical AI model for real-time intention understanding; dynamically retrieves the vertical domain knowledge related to the current interaction scenario, extracts the domain knowledge features, and inputs them into the vertical AI model for knowledge enhancement; utilizes the decision-making mechanism of the vertical AI model, integrates the understood user intentions and related domain knowledge, and outputs personalized service decisions; applies personalized service decisions to the Metaverse platform, and improves the user experience through intelligent service forms such as recommendation, search, question and answer, and interactive assistance.

[0178] The trained and optimized vertical AI model is deployed and applied to provide intelligent personalized services for the Metaverse platform. When users interact in different vertical scenarios, their behavior data is collected in real time and the intent features are extracted and input into the model for intent understanding. Since the model has been pre-trained on a heterogeneous knowledge graph, it can make full use of the association structure between intent nodes and accurately characterize the dynamic intent of users from their behavior trajectories. At the same time, for the user's current interaction scenario, the relevant domain nodes in the knowledge graph are retrieved in real time to capture the concepts, entities and their relationships related to intent understanding and decision-making, form knowledge features and input into the model. Online retrieval ensures that the domain knowledge used by the model matches the actual scenario requirements, thereby improving the accuracy and timeliness of personalized services. The model uses the pre-trained attention mechanism to adaptively adjust the contribution of different knowledge information to intent understanding, and combines the temporal characteristics of intent to generate personalized service decisions for current interaction needs. The decision results can support different forms of intelligent services, such as personalized recommendations, intelligent search, knowledge questions and answers, and interactive assistance. The Metaverse platform implements specific services based on service decisions and feeds back the service effects to the model to form a data closed loop.

[0179] Example 2

[0180] Based on Example 1, this example provides a multi-dimensional fusion metaverse and vertical AI model collaborative innovation method, such as Figure 2 As shown, including:

[0181] Step S1000, collecting user interaction behavior data, extracting a first data set from the interaction behavior data; analyzing the first data set, identifying the user's explicit intention and implicit intention, and generating a second data set; based on the second data set, constructing a user-centered interaction intention knowledge graph;

[0182] Furthermore, step S1000 includes:

[0183] Step S1100, collecting user interaction behavior data, and extracting a first data set from the interaction behavior data;

[0184] Furthermore, if Figure 3 As shown, step S1100 includes:

[0185] Step S1110, recording the user's interactive operations including clicks, slides and voice inputs in real time to form an original interaction log;

[0186] Step S1120, extracting key interaction fields including user ID, interaction object ID, interaction timestamp and interaction type from the original interaction log to form a structured interaction behavior record;

[0187] Step S1130 , aggregating the interaction behavior records according to the user ID to generate a first data set, wherein the first data set includes the user's interaction behavior sequence.

[0188] Step S1200, analyzing the first data set, identifying the user's explicit intention and implicit intention, and generating a second data set;

[0189] Further, step S1200 includes:

[0190] Step S1210, encoding the interaction behavior sequence in the first data set, and extracting the semantic feature vector of the interaction behavior sequence;

[0191] Further, step S1210 includes:

[0192] Step S1211: the interaction behavior sequence [(object 1, operation 1), (object 2, operation 2), ..., (object n, operation n)] contains n interaction events (object, operation), and each interaction event is mapped into a low-dimensional dense vector as an embedded representation of the interaction event;

[0193] Step S1212, combining the embedded representations of n interaction events into an event embedding sequence;

[0194] Step S1213, obtaining a semantic feature vector of the interaction behavior sequence according to the event embedding sequence.

[0195] Step S1220, clustering the semantic feature vectors, identifying explicit intents according to the clustering results, and associating them with predefined intent types to obtain explicit intent types;

[0196] Further, step S1220 includes:

[0197] Step S1221, calculating the similarity matrix between semantic feature vectors;

[0198] Step S1222, constructing a minimum spanning tree according to the similarity matrix;

[0199] Step S1223, setting a distance threshold on the minimum spanning tree , according to the distance threshold Get the clustering of semantic feature vectors;

[0200] Step S1224, matching the cluster of each semantic feature vector with the predefined intent type to obtain the explicit intent type of the cluster.

[0201] Step S1230, establishing an intention evolution model, obtaining a hidden state sequence of the user interaction behavior sequence through the intention evolution model, and obtaining the user's implicit intention and implicit intention type according to the hidden state sequence;

[0202] Step S1240, hierarchically organizing the identified explicit intentions and implicit intentions, and generating a second data set in combination with the time sequence of the intentions, wherein each data sample in the second data set includes an intention type, an intention level, and an intention time sequence;

[0203] Furthermore, if Figure 4 As shown, step S1240 includes:

[0204] Step S1241, constructing a hierarchical intent type system according to the semantic relationship between intent types;

[0205] Step S1242, mapping each user interaction behavior sequence to the intention type system, and assigning a corresponding intention type label to each interaction event in the interaction behavior sequence;

[0206] Step S1243, recording the timestamp of each intent type to obtain the time sequence of the intent sequence;

[0207] Step S1244, generating a second data set according to the time order of the intention sequence.

[0208] Step S1300, constructing a user-centric interaction intention knowledge graph based on the second data set;

[0209] Furthermore, step S1300 includes:

[0210] Step S1310, mapping the intent type, intent hierarchy, and intent sequence in the second data set to nodes and edges in the knowledge graph;

[0211] Further, step S1310 includes:

[0212] Step S1311, mapping the intent type to an intent type node in the knowledge graph;

[0213] Step S1312, mapping the intent instances at different time points of the intent time series in the second data set to intent instance nodes under corresponding type nodes;

[0214] Step S1313, constructing intent timing edges between intent instance nodes based on intent timing;

[0215] Step S1314, mapping each user to a user node in the knowledge graph;

[0216] Step S1315, based on the hierarchical relationship between intent types, establish a subclass relationship edge between corresponding intent type nodes; and establish an instance relationship edge between the intent instance node and the corresponding intent type node.

[0217] Step S1320, associating a series of intent instance nodes of the same user at different time points with the user node of the user, to form a star topology subgraph with the user node as the center and the intent instance nodes surrounding it;

[0218] Step S1330, adding attribute information to the nodes and edges in the knowledge graph to obtain the final interaction intention knowledge graph;

[0219] Further, step S1330 includes:

[0220] Step S1331, for the intent type node, add statistical attributes including word frequency and TF-IDF;

[0221] Step S1332, for the intent instance node, add context attributes including timestamp, geographic location, and emotional tendency;

[0222] Step S1333: for the intention time sequence edge, add dynamic attributes including transition probability and time decay coefficient;

[0223] Step S1334: for the user node, associate the user's demographic attributes and preference attributes.

[0224] Step S2000, identifying key vertical domain knowledge of different metaverse application scenarios, constructing a vertical domain knowledge graph based on the vertical domain knowledge, and fusing the vertical domain knowledge graph with the interaction intention knowledge graph to obtain a fused heterogeneous knowledge graph;

[0225] Furthermore, step S2000 includes:

[0226] Step S2100, identifying key vertical domain knowledge for different Metaverse application scenarios;

[0227] Furthermore, step S2100 includes:

[0228] Step S2110, determining the key vertical fields of Metaverse application scenarios;

[0229] Step S2120, for each vertical field, collecting structured knowledge including field ontology, knowledge base and industry dictionary and unstructured knowledge including field documents and user comments;

[0230] Step S2130, constructing vertical domain knowledge from structured knowledge and unstructured knowledge.

[0231] Step S2200, constructing a vertical domain knowledge graph based on the vertical domain knowledge, and fusing the vertical domain knowledge graph with the interaction intention knowledge graph to obtain a fused heterogeneous knowledge graph;

[0232] Further, step S2200 includes:

[0233] Step S2210, constructing a vertical domain knowledge graph based on vertical domain knowledge;

[0234] Further, step S2210 includes:

[0235] Step S2211, defining the core concepts, entities, entity types, and attributes of the vertical field, and mapping the core concepts, entities, entity types, and attributes of the vertical field to nodes in the vertical field knowledge graph;

[0236] Step S2212, mapping the relationships between entities in the vertical field into edges in the knowledge graph;

[0237] Step S2213, based on the hierarchical classification system of the vertical field, establish subclass relationship edges between corresponding entity type nodes;

[0238] Step S2214, establishing logical regular edges between related entity nodes;

[0239] Step S2215, map the events in the vertical field into event nodes in the vertical field knowledge graph, and establish associated edges with related entity nodes.

[0240] Step S2220, using a graph neural network to learn dense embedding representations of nodes and edges in a vertical domain knowledge graph, and generating an initial graph embedding vector including a node embedding vector and an edge embedding vector; the node embedding vector includes an entity node embedding vector;

[0241] Step S2230, propagating the node embedding vectors in the initial graph embedding vector through the graph convolutional network, so that entities with similar semantics or structures in the graph obtain similar node embedding vectors, and updating the node embedding vectors in the initial graph embedding vector with the similar node embedding vectors to obtain an optimized graph embedding vector;

[0242] Step S2240, associating the optimized graph embedding vector of the vertical field with the nodes in the interactive intent knowledge graph to obtain a fused heterogeneous knowledge graph; the nodes in the interactive intent knowledge graph include user nodes and intent instance nodes;

[0243] Further, step S2240 includes:

[0244] Step S2241, obtaining the node embedding vector of the interaction intention knowledge graph;

[0245] Step S2242, calculate the similarity between the node embedding vector of the interaction intention knowledge graph and the node embedding vector of the vertical field knowledge graph, and set the similarity of the node embedding vector to be greater than the similarity threshold The interactive intention knowledge graph nodes and the nodes of the vertical domain knowledge graph establish semantic associations between cross-graph nodes;

[0246] Step S2243, using the semantic association between cross-graph nodes established in step S2242, construct a node alignment loss function; based on the node alignment loss function and the structure preservation loss of a single graph, establish a comprehensive optimization function, and obtain the fused heterogeneous knowledge graph by minimizing the comprehensive loss function; the structure preservation loss of the single graph includes the structure preservation loss function of the interactive intent knowledge graph and the structure preservation loss function of the vertical domain knowledge graph.

[0247] Step S3000: construct a vertical AI model based on the fused heterogeneous knowledge graph, and apply the vertical AI model to understand and predict user intentions to achieve intelligent personalized services;

[0248] Furthermore, step S3000 includes:

[0249] Step S3100, extracting user intention features and domain knowledge features according to the fused heterogeneous knowledge graph, combining the user intention features and the domain knowledge features, and constructing a training data set for the vertical AI model;

[0250] Step S3200, constructing a vertical AI model, and using a training data set to perform end-to-end training and optimization on the vertical AI model; the vertical AI model is a deep learning model having a graph neural network encoder and an attention mechanism decoder structure;

[0251] Further, step S3200 includes:

[0252] Step S3210, using the user intention features in the training data set to input the graph neural network encoder to obtain an encoded user intention representation;

[0253] Step S3220, using the domain knowledge features in the training data set to input the graph neural network encoder to obtain the encoded domain knowledge representation;

[0254] Step S3230, applying the attention mechanism on the encoded user intention representation and domain knowledge representation, calculating the intention-knowledge interaction weight, and obtaining the intention-knowledge combination representation;

[0255] Step S3240, input the intention-knowledge combination representation into the decoder, decode it to generate a personalized service decision, and calculate the loss function of the decision;

[0256] Step S3250, using the stochastic gradient descent algorithm to minimize the loss function of the decision, and perform end-to-end training and optimization on the parameters of the vertical AI model;

[0257] Step S3300, apply the trained and optimized vertical AI model to realize intelligent personalized services for different vertical scenarios of the Metaverse and enhance the user interaction experience.

[0258] Example 3

[0259] This embodiment discloses an electronic device 500, such as Figure 5 As shown, the electronic device 500 may include one or more processors and one or more memories. The memories store computer-readable codes, which, when executed by one or more processors, can execute a multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform as described above.

[0260] The method or system according to the embodiment of the present application can also be Figure 5 The electronic device architecture shown in FIG. Figure 5As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, a read-only memory (ROM) 503, a random access memory (RAM) 504, a communication port 505 connected to a network, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as the ROM 503 or the hard disk 507, may store a multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform provided in the present application. A multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform may, for example, include: collecting user interaction behavior data, extracting a first data set from the interaction behavior data; analyzing the first data set, identifying the user's explicit intentions and implicit intentions, and generating a second data set; based on the second data set, constructing a user-centered interaction intention knowledge graph; identifying key vertical domain knowledge for different metaverse application scenarios, constructing a vertical domain knowledge graph based on the vertical domain knowledge, and fusing the vertical domain knowledge graph with the interaction intention knowledge graph to obtain a fused heterogeneous knowledge graph; based on the fused heterogeneous knowledge graph, constructing a vertical AI model, and applying the vertical AI model to understand and predict user intentions to achieve intelligent personalized services.

[0261] Furthermore, the electronic device 500 may also include a user interface 508. Figure 5 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 5 One or more components of an electronic device are shown.

[0262] Example 4

[0263] This embodiment discloses a computer-readable storage medium 600, such as Figure 6 As shown, computer-readable instructions are stored on the computer-readable storage medium 600. When the computer-readable instructions are executed by the processor, a multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to the embodiment of the present application described with reference to the above figures can be executed. The storage medium 600 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache). Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0264] In addition, according to the implementation of the present application, the process described in the above reference flowchart can be implemented as a computer software program. For example, the present application provides a non-temporary machine-readable storage medium, which stores machine-readable instructions, and the machine-readable instructions can be executed by a processor to execute instructions corresponding to the method steps provided in the present application, such as: collecting user interaction behavior data, extracting a first data set from the interaction behavior data; analyzing the first data set, identifying the user's explicit intentions and implicit intentions, and generating a second data set; based on the second data set, constructing a user-centered interaction intention knowledge graph; identifying key vertical domain knowledge for different metaverse application scenarios, constructing a vertical domain knowledge graph based on the vertical domain knowledge, and fusing the vertical domain knowledge graph with the interaction intention knowledge graph to obtain a fused heterogeneous knowledge graph; based on the fused heterogeneous knowledge graph, constructing a vertical AI model, and applying the vertical AI model to understand and predict user intentions to achieve intelligent personalized services. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0265] The methods, apparatuses, and devices of the present application may be implemented in many ways. For example, the methods, apparatuses, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0266] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0267] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-dimensional fusion of metaverse and vertical AI model collaborative innovation platform, characterized by: The platform includes: a data generation module, an interaction intention knowledge graph construction module, a vertical field knowledge graph construction module, a knowledge graph fusion module and an intelligent interaction module; The data generation module is used to collect the user's interactive behavior data, extract a first data set from the interactive behavior data; analyze the first data set, identify the user's explicit intention and implicit intention, and generate a second data set; The interaction intention knowledge graph construction module is used to construct a user-centric interaction intention knowledge graph based on the second data set; The vertical domain knowledge graph construction module is used to identify key vertical domain knowledge for different metaverse application scenarios and construct a vertical domain knowledge graph based on the vertical domain knowledge; The knowledge graph fusion module is used to fuse the vertical domain knowledge graph with the interactive intention knowledge graph to obtain a fused heterogeneous knowledge graph; The intelligent interaction module is used to build a vertical AI model based on the fused heterogeneous knowledge graph, apply the vertical AI model to understand and predict user intentions, and realize intelligent personalized services.

2. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 1 is characterized in that: The data generation module includes a first data generation unit and a second data generation unit; The first data generation unit is used to record the user's interactive operations including clicks, slides and voice inputs in real time to form an original interaction log; extract key interaction fields including user ID, interaction object ID, interaction timestamp and interaction type from the original interaction log to form a structured interaction behavior record; aggregate the interaction behavior records according to user ID to generate a first data set, which includes the user's interaction behavior sequence.

3. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 2 is characterized in that: The second data generating unit is used to encode the interaction behavior sequence in the first data set and extract the semantic feature vector of the interaction behavior sequence; Cluster the semantic feature vectors, identify explicit intents based on the clustering results, and associate them with predefined intent types to obtain explicit intent types; establish an intent evolution model, obtain the hidden state sequence of the user interaction behavior sequence through the intent evolution model, and obtain the user's implicit intent and implicit intent type based on the hidden state sequence; Organizing the identified explicit intentions and implicit intentions in a hierarchical manner, and combining the time sequence of the intentions to generate a second data set, wherein each data sample in the second data set includes an intention type, an intention level, and an intention time sequence; The second data generating unit includes a data encoding unit, an explicit intention recognition unit, an implicit intention recognition unit and a data combining unit.

4. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 3 is characterized in that: The data encoding unit is used to map each interaction event in the interaction behavior sequence into a low-dimensional dense vector as an embedded representation of the interaction event, wherein the interaction behavior sequence includes n interaction events; combine the embedded representations of the n interaction events into an event embedding sequence; and obtain a semantic feature vector of the interaction behavior sequence according to the event embedding sequence; The explicit intention recognition unit is used to calculate the similarity matrix between semantic feature vectors; Construct a minimum spanning tree based on the similarity matrix; set a distance threshold on the minimum spanning tree , according to the distance threshold Obtaining a cluster of semantic feature vectors; matching each cluster of semantic feature vectors with a predefined intent type to obtain an explicit intent type of the cluster; The implicit intention recognition unit is used to establish an intention evolution model, obtain a hidden state sequence of the user interaction behavior sequence through the intention evolution model, and obtain the user's implicit intention and implicit intention type according to the hidden state sequence; The data combination unit is used to construct a hierarchical intent type system according to the semantic relationship between intent types; Map each user interaction behavior sequence to the intent type system, and assign a corresponding intent type label to each interaction event in the interaction behavior sequence; Record the timestamp of each intent type to obtain the time sequence of the intent sequence; A second data set is generated according to the temporal order of the intent sequence.

5. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 1 is characterized in that: The interactive intention knowledge graph construction module includes a graph mapping unit, a node association unit and an attribute attachment unit; The graph mapping unit is used to map the intent type, intent hierarchy, and intent sequence in the second data set into nodes and edges in the knowledge graph; The node association unit is used to associate a series of intent instance nodes of the same user at different time points with the user node of the user, forming a star topology subgraph with the user node as the center and the intent instance nodes surrounding it; The attribute adding unit is used to add attribute information to the nodes and edges in the knowledge graph to obtain the final interaction intention knowledge graph.

6. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 5 is characterized in that: The mapping of intent types, intent hierarchies, and intent sequences in the second data set to nodes and edges in the knowledge graph includes: mapping intent types to intent type nodes in the knowledge graph; mapping intent instances at different time points in the intent sequence in the second data set to intent instance nodes under corresponding type nodes; constructing intent sequence edges between intent instance nodes based on intent sequence; mapping each user to a user node in the knowledge graph; establishing subclass relationship edges between corresponding intent type nodes based on the hierarchical relationship between intent types; and establishing instance relationship edges between intent instance nodes and corresponding intent type nodes.

7. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 1 is characterized in that: The vertical field knowledge graph construction module includes a knowledge identification unit and a graph construction unit; The knowledge identification unit is used to determine the key vertical fields of the metaverse application scenarios; for each vertical field, it collects structured knowledge including domain ontology, knowledge base and industry dictionary and unstructured knowledge including domain documents and user comments; and constructs vertical domain knowledge from structured knowledge and unstructured knowledge.

8. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 1, characterized in that: The knowledge graph fusion module includes an embedding vector generation unit, an embedding vector optimization unit and a graph fusion unit; The embedding vector generating unit is used to learn the dense embedding representation of nodes and edges in the vertical domain knowledge graph using a graph neural network, and generate an initial graph embedding vector including a node embedding vector and an edge embedding vector; the node embedding vector includes an entity node embedding vector; The embedding vector optimization unit is used to propagate the node embedding vectors in the initial graph embedding vector through the graph convolution network, so that entities with similar semantics or structures in the graph obtain similar node embedding vectors, and use the similar node embedding vectors to update the node embedding vectors in the initial graph embedding vector to obtain an optimized graph embedding vector; The graph fusion unit is used to associate the optimized graph embedding vector of the vertical field with the nodes in the interactive intent knowledge graph to obtain a fused heterogeneous knowledge graph; the nodes in the interactive intent knowledge graph include user nodes and intent instance nodes.

9. The multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform according to claim 8, characterized in that: The method of associating the optimized graph embedding vector of the vertical field with the node in the interactive intention knowledge graph includes: obtaining the node embedding vector of the interactive intention knowledge graph; calculating the similarity between the node embedding vector of the interactive intention knowledge graph and the node embedding vector of the vertical field knowledge graph, and calculating the similarity of the node embedding vector greater than the similarity threshold. The nodes of the interactive intention knowledge graph and the nodes of the vertical domain knowledge graph establish semantic associations between cross-graph nodes; the semantic associations between cross-graph nodes are used to construct a node alignment loss function; according to the node alignment loss function and the structure preservation loss of a single graph, a comprehensive optimization function is established, and the fused heterogeneous knowledge graph is obtained by minimizing the comprehensive loss function; the structure preservation loss of the single graph includes the structure preservation loss function of the interactive intention knowledge graph and the structure preservation loss function of the vertical domain knowledge graph.

10. A multi-dimensional fusion metaverse and vertical AI model collaborative innovation method, which is based on the multi-dimensional fusion metaverse and vertical AI model collaborative innovation platform described in any one of claims 1-9, characterized in that: The method comprises: Collecting user interaction behavior data, extracting a first data set from the interaction behavior data; analyzing the first data set, identifying the user's explicit intention and implicit intention, and generating a second data set; constructing a user-centered interaction intention knowledge graph based on the second data set; Identify key vertical domain knowledge for different metaverse application scenarios, build vertical domain knowledge graphs based on vertical domain knowledge, and fuse the vertical domain knowledge graphs with the interaction intent knowledge graphs to obtain fused heterogeneous knowledge graphs; Based on the fused heterogeneous knowledge graph, a vertical AI model is constructed, and the vertical AI model is applied to understand and predict user intentions to achieve intelligent personalized services.

11. An electronic device comprising a memory, a central processing unit, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the multi-dimensional fusion metaverse and vertical AI model collaborative innovation method described in claim 10 is implemented.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed, implements the multi-dimensional fusion metaverse and vertical AI model collaborative innovation method described in claim 10.

Images