Method for generating user digital avatar behavior patterns based on metaverse

Abstract

The application discloses a user digital avatar behavior mode generation method based on a meta universe, relates to the technical field of digital avatar learning, processes collected user social data, meta universe behavior data and behavior context to obtain user behavior event data; the user behavior event data is processed by a multi-channel text feature model to obtain a text feature vector, a text classification probability is obtained through multi-channel task processing, a mixed kernel embedding learning model is constructed, the text classification probability and the behavior context are mapped to obtain user behavior mapping data, and finally, the user behavior mapping data and the current meta universe environment are analyzed to obtain user behavior parameters, so that personalized autonomous behavior can be maintained when the user is offline, and deep behavior preferences and decision logic of the user can be effectively and timely analyzed.

Description

Technical Field

[0001] This invention relates to the field of digital avatar learning technology, and in particular to a method for generating user digital avatar behavior patterns based on the metaverse. Background Technology

[0002] With the rapid development of metaverse technology, digital avatars, as the core agents and identity carriers of users in the virtual world, have become key factors in determining user experience and immersion in terms of the intelligence, authenticity, and personalization of their behavior.

[0003] Currently, the generation of digital avatar behaviors has significant limitations. Methods based on predefined rules and animation libraries, which involve manually writing behavior rules and binding them to preset animation libraries to achieve avatar behavior, are simple to implement and respond quickly, but their behavior patterns are rigid, lack scalability, fail to reflect users' individual habits and complex intentions, and are even more difficult to adapt to open and dynamic metaverse environments, resulting in mechanical and unnatural avatar behavior. Secondly, methods based on real-time user control and mapping offer greater flexibility, mapping users' real-time actions and expressions to digital avatars through motion capture, voice commands, or direct control interfaces; however, this approach… The current system heavily relies on users' continuous online presence and active control, failing to address the issue of avatars disconnecting when users are offline. When a user leaves, their digital avatar often becomes static or executes simple looping animations, disrupting the continuity of the metaverse social scene and preventing users from using avatars for asynchronous social interaction or task progression. Furthermore, traditional machine learning and data analysis methods attempt to analyze users' chat logs and movement trajectories from historical behavioral data, using clustering or classification algorithms to summarize a few rough behavioral preferences. However, these methods typically only process single-modal data, lacking a deep fusion and understanding of multimodal social data and rich behavioral context. The generated models often remain at a superficial level, unable to uncover deep, fine-grained behavioral motivations and emotional tendencies, leading to a disconnect between avatar behavior and the user's true personality, and insufficient generalization ability in complex scenarios. Summary of the Invention

[0004] The technical problem solved by this invention is that digital clones cannot maintain personalized autonomous behavior when users are offline, and traditional methods are highly dependent on real-time user operations, making it impossible to adaptively analyze users' deep-seated behavioral preferences and decision-making logic.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for generating user digital avatar behavior patterns based on the metaverse, comprising the following steps:

[0006] Step S1: Process the collected user social data, metaverse behavior data, and behavior context to obtain user behavior event data;

[0007] Step S2: Process the user behavior event data using a multi-channel text feature model to obtain text feature vectors, and obtain text classification probabilities through multi-channel task processing.

[0008] Step S3: Construct a hybrid kernel embedding learning model to map text classification probabilities and behavioral context to obtain user behavior mapping data;

[0009] Step S4: Analyze the user behavior mapping data and the current metaverse environment to obtain user behavior parameters.

[0010] Preferably, step S1 includes the following sub-steps:

[0011] Step S11: Obtain user social data, metaverse behavior data, and behavior context through the data collection SDK built into the metaverse client. The user social data includes text data, image data, and interactive behavior data. The metaverse behavior data includes micro-expressions, action data, and social behavior data. The behavior context includes timestamps, locations, scenes, activity types, and interacting users.

[0012] Step S12: Construct a social network structure using user social data. The social network structure includes a friend attention network and a community participation network. The friend attention network uses the user as the main node and each friend user as the edge node. The friend edge weight is used to connect the user with each friend user. The friend edge weight is the proportion of the user's chat time with a single friend user to the user's total chat time within a unit period. The community participation network uses the user as the main node and each community as the edge node. The community edge weight is used to connect the user with each community. The community edge weight is the proportion of the user's chat time in a single community to the user's chat time in all communities within a unit period.

[0013] Step S13: Match user social data, metaverse behavior data, and behavior context one-to-one according to the same timestamp to obtain metaverse interaction data. Set a sliding time window and divide the metaverse interaction data into user behavior event data according to the sliding time window.

[0014] Preferably, step S2 includes the following sub-steps:

[0015] Step S21: Use a multi-channel text feature model to perform self-attention pooling on the text data in the user behavior event data to obtain text feature vectors;

[0016] Step S22: Map the text feature vectors to a shared low-dimensional semantic space to obtain a low-dimensional shared semantic vector. The calculation expression for the low-dimensional shared semantic vector is as follows:

[0017] ;

[0018] in, For activation function, For layer normalization, For low-dimensional shared semantic vectors, and For learning parameters, For text feature vectors;

[0019] Step S23: The output text feature vector is classified using a multi-head task parallel channel to obtain the text classification probability, which includes the probability of utilitarian imbalance, the probability of emotional imbalance, and the probability of normal interaction.

[0020] Preferably, step S21 further includes the following sub-steps:

[0021] Step S211: The text data that has been divided in the behavioral event data is concatenated in chronological order to construct a long text sequence. At the same time, the interactive behavioral data and metaverse behavioral data that occur at the corresponding time points are converted into predefined behavioral tags and inserted into the time positions corresponding to the long text sequence to generate a long sequence. The long sequence includes text data and metaverse behavioral data. A session unit tag is inserted before each text message in the long text sequence. A sliding time window is set. An event start tag is inserted at the beginning of each sliding time window. A user anchor tag is inserted at the first appearance position of the analyzed user to obtain a metaverse interaction long sequence.

[0022] Step S212: Construct a multi-channel embedding using the meta-universe interaction long sequence to obtain the input representation vector. The multi-channel embedding includes word action embedding, absolute position embedding, time interval embedding, role relationship embedding, and social structure embedding.

[0023] The word-behavior embedding is used to establish word embedding matrices and behavior tag embedding matrices for text words and behavior tags, respectively;

[0024] The absolute position embedding is used to encode the position index based on the markers in the metaverse interaction long sequence;

[0025] The time interval embedding is used to perform segmented encoding based on the time difference between the behavior time corresponding to the long sequence of metaverse interactions and the starting time of the sliding window;

[0026] The role relationship embedding is used to distinguish different conversational participation roles of target users, their own digital clones, user friends, friends in the same community, and strangers.

[0027] The social structure embedding is used to map friend edge weights and community edge weights into low-dimensional vectors based on the friend attention network and community participation network.

[0028] Step S213: Input the input representation vector into the precoding layer to obtain the initial hidden representation. Calculate the global importance score for each tag based on the initial hidden representation. The expression for calculating the global importance score is:

[0029] ;

[0030] in, For the Sigmoid function, This is the initial hidden representation. and For learnable parameters, A global importance score is set with a preset score threshold. When the global importance score is greater than the score threshold, the current corresponding label is set as the global attention label, and the remaining labels are treated with local sliding window attention to generate a sparse attention structure.

[0031] Step S214: Longformer encoding and gated self-attention pooling are used to perform deep representation processing on the initial hidden representation to obtain the text feature vector.

[0032] Preferably, step S3 includes the following sub-steps:

[0033] Step S31: Standardize and encode the text classification probability and behavioral context to obtain the text classification vector and behavioral context vector;

[0034] Step S32: Using a hybrid kernel embedding learning model, the text classification vector and the behavior context vector are fused and analyzed to output user behavior mapping data. The user behavior mapping data includes user behavior parameters and context labels. The user behavior parameters include social initiative parameters, emotional expression intensity parameters, topic preference parameters, response speed parameters, response length parameters, emotional matching degree parameters, and task engagement degree parameters.

[0035] Preferably, the processing logic for fusing and analyzing text classification vectors and behavioral context vectors using a hybrid kernel embedding learning model is as follows:

[0036] The text classification vector is mapped to a high-dimensional semantic space using a deep variational Gaussian kernel to obtain semantic distribution features. The behavioral context vector is constructed using a spectral graph convolutional network to create a structural representation, which is then mapped using a Gaussian process to obtain behavioral structural features. Finally, a hybrid semantic kernel embedding is used to fuse the semantic distribution features and behavioral structural features to generate hybrid semantic features. The calculation expression for the hybrid semantic kernel embedding is as follows:

[0037] ;

[0038] in, It is a hybrid semantic feature. Semantic distribution features As behavioral structural features, For fusion weights;

[0039] The hybrid semantic features are input into the behavior parameter decoder, which outputs user behavior parameters. The user behavior parameters are classified according to the triple self-supervised scenario optimization, and the corresponding scenario labels are obtained. The user behavior parameters and scenario labels are saved as user behavior mapping data.

[0040] Preferably, the processing logic for classifying user behavior parameters based on triple self-supervised scenario optimization and obtaining corresponding scenario labels is as follows:

[0041] Calculate the soft assignment probability of the mixed semantic features to the corresponding context labels. The expression for calculating the soft assignment probability is as follows:

[0042]

[0043] in, Let i be the hybrid semantic kernel embedding vector of the i-th sample. Let the vector centered at the c-th row represent a typical behavioral pattern. Degrees of freedom allocated to behavior The soft-assignment probability of the mixed semantic features of the i-th sample corresponding to the c-th row is given.

[0044] The target distribution is calculated for the soft allocation probability, and the expression for the target distribution is as follows:

[0045]

[0046] in, For the target distribution, The soft-assignment probabilities of all samples i are summed; the output scenario category labels are optimized by using an adversarial loss function and gradient descent, and user behavior parameters are mapped one-to-one with the corresponding scenario category labels and linked to all behavior centers. Save as user behavior mapping data.

[0047] Preferably, step S4 includes the following sub-steps:

[0048] Step S41: Obtain the current environment data of the user when offline from the metaverse environment interface, preprocess the current environment data, and generate the current environment feature vector;

[0049] Step S42: Using the scenario category matching algorithm, calculate the cosine similarity between the current environment feature vector and the center of each row in the user behavior mapping data, and select the behavior center with the highest cosine similarity as the current scenario category label.

[0050] Step S43: Based on the current scenario category label, retrieve all user behavior parameters belonging to the scenario category label from the user behavior mapping data, and perform a weighted average of the user behavior parameters to obtain the user behavior parameters in the current environment.

[0051] Preferably, the processing logic of the scenario category matching algorithm is as follows:

[0052] The current environment data is preprocessed to generate the cosine similarity between the current environment feature vector and the feature vector corresponding to each scenario category label. The scenario category with the highest cosine similarity is selected as the current scenario category label.

[0053] Preferably, the scenario category labels include:

[0054] Collaborative work, social entertainment, exploratory learning, and leisure observation.

[0055] The beneficial effects of this invention are as follows: First, by using a multi-channel text feature model and sequential multimodal embedding, it innovatively unifies the encoding of discrete actions and continuous text, solving the problem of heterogeneous data fusion and laying the foundation for a deep understanding of user interaction intentions. Second, by employing a hybrid kernel embedding learning model, semantic distribution features are extracted through a deep variational Gaussian kernel, and structured features are extracted through a spectral graph convolutional network and a Gaussian process. Weighted fusion is then performed in the regenerating kernel Hilbert space, fundamentally overcoming the information loss or conflict problems that traditional splicing or simple attention mechanisms encounter when fusing heterogeneous features. Finally, the triple self-supervised scenario optimization mechanism can automatically mine and solidify typical behavior patterns from user historical data without manual annotation, greatly reducing the dependence on large-scale labeled data and improving the model's adaptability and interpretability. Through the design of step S4, the computationally complex hybrid kernel model training and scenario clustering are placed offline. During online generation, only efficient cosine similarity calculation is needed to retrieve and fuse parameters from the established behavior mapping library, thereby achieving millisecond-level response and meeting the needs of real-time interaction in the metaverse. Attached Figure Description

[0056] Figure 1 A flowchart illustrating the steps of a method for generating user digital clone behavior patterns based on a metaverse, as provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0058] Example, refer to Figure 1It provides a method for generating user digital avatar behavior patterns based on the metaverse, including the following steps:

[0059] Step S1: Process the collected user social data, metaverse behavior data, and behavior context to obtain user behavior event data;

[0060] Step S2: Process the user behavior event data using a multi-channel text feature model to obtain text feature vectors, and obtain text classification probabilities through multi-channel task processing.

[0061] Step S3: Construct a hybrid kernel embedding learning model to map text classification probabilities and behavioral context to obtain user behavior mapping data;

[0062] Step S4: Analyze the user behavior mapping data and the current metaverse environment to obtain user behavior parameters.

[0063] By innovatively unifying discrete actions and continuous text through a multi-channel text feature model and sequential multimodal embedding, the problem of heterogeneous data fusion is solved, laying the foundation for a deep understanding of user interaction intentions. Secondly, a hybrid kernel embedding learning model is adopted, which extracts semantic distribution features through a deep variational Gaussian kernel and extracts structured features through a spectral graph convolutional network and a Gaussian process. Weighted fusion is then performed in the regenerating kernel Hilbert space, fundamentally overcoming the information loss or conflict problems of traditional splicing or simple attention mechanisms when fusing heterogeneous features. Finally, a triple self-supervised scenario optimization mechanism can automatically mine and solidify typical behavior patterns from user historical data without manual annotation, greatly reducing the dependence on large-scale labeled data and improving the model's adaptability and interpretability. Through the design of step S4, the computationally complex hybrid kernel model training and scenario clustering are placed offline. During online generation, only efficient cosine similarity calculation is needed to retrieve and fuse parameters from the established behavior mapping library, thereby achieving millisecond-level response and meeting the needs of real-time interaction in the metaverse.

[0064] Step S1 includes the following sub-steps:

[0065] Step S11: Obtain user social data, metaverse behavior data and behavior context through the data collection SDK built into the metaverse client. User social data includes text data, image data and interactive behavior data. Metaverse behavior data includes micro-expressions, action data and social behavior data. Behavior context includes timestamps, locations, scenes, activity types and interacting users.

[0066] Step S12: Construct a social network structure using user social data. The social network structure includes a friend attention network and a community participation network. The friend attention network uses the user as the main node and each friend user as the edge node. The friend edge weight is used to connect the user with each friend user. The friend edge weight is the proportion of the user's chat time with a single friend user to the user's total chat time within a unit period. The community participation network uses the user as the main node and each community as the edge node. The community edge weight is used to connect the user with each community. The community edge weight is the proportion of the user's chat time in a single community to the user's chat time in all communities within a unit period.

[0067] Step S13: Match user social data, metaverse behavior data, and behavior context one-to-one according to the same timestamp to obtain metaverse interaction data. Set a sliding time window and divide the metaverse interaction data into user behavior event data according to the sliding time window.

[0068] Text data includes posts, comments, private messages, and group chat records; image data includes selfies, screenshots, and emojis; interactive behavior data includes likes, reposts, comments, @ mentions, blocking, unfollowing, blocking keywords, and muting; metaverse behavior data includes micro-expressions, mapped to digital human expressions; action data includes body postures, gestures, approaching / moving away from someone, observing / passing by, etc.; social behavior data includes the number of times, the target audience, and the topic of actively initiating conversations; passive response rate, response latency, and conversation depth; and the frequency and target audience of teaming up / joining / leaving a room. The unit period is set to one week. For example, if user A chats privately with friend B for a total of 5 hours and with all friends for a total of 40 hours in a week, then the edge weight between user A and friend B is 0.125.

[0069] In step S13, the sliding time window is set to 5 minutes, and the continuous interactive data stream is divided into independent user behavior event data segments according to the sliding time window. Each user behavior event data packet contains all aligned multimodal interaction records within a time period. In this embodiment, by aligning multi-source data, the continuous behavior stream is transformed into a limited number of independent analysis units through slicing, which facilitates subsequent batch processing.

[0070] Step S2 includes the following sub-steps:

[0071] Step S21: Use a multi-channel text feature model to perform self-attention pooling on the text data in the user behavior event data to obtain text feature vectors;

[0072] Step S22: Map the text feature vectors to a shared low-dimensional semantic space to obtain a low-dimensional shared semantic vector. The calculation expression for the low-dimensional shared semantic vector is:

[0073] ;

[0074] in, For activation function, For layer normalization, For low-dimensional shared semantic vectors, and For learning parameters, For text feature vectors;

[0075] Step S23: The output text feature vector is classified using a multi-head task parallel channel to obtain the text classification probability, which includes the probability of utilitarian imbalance, the probability of emotional imbalance, and the probability of normal interaction.

[0076] Using the user behavior event data from step S13, step S21 is executed. This user behavior event data is divided into data blocks according to a sliding time window of five minutes, including all time-aligned text messages, behavior tags, and multi-channel embeddings within the window. Key features are extracted from massive text through a self-attention mechanism to provide high-quality input for subsequent classification and to prepare for semantic mapping in step S22. Then, the text feature vectors are mapped to a low-dimensional semantic space to unify the semantic representation of different behavior events. By further compressing the abstract features, the curse of dimensionality is reduced. Step S23 adopts multi-head parallel classification to identify multiple interaction features simultaneously, including utilitarian imbalance, emotional fit, and normal interaction, to ensure knowledge sharing between tasks.

[0077] Step S21 also includes the following sub-steps:

[0078] Step S211: The text data that has been divided in the behavioral event data is concatenated in chronological order to construct a long text sequence. At the same time, the interactive behavioral data and metaverse behavioral data that occur at the corresponding time points are converted into predefined behavioral tags and inserted into the time positions corresponding to the long text sequence to generate a long sequence. The long sequence includes text data and metaverse behavioral data. A session unit tag is inserted before each text message in the long text sequence. A sliding time window is set, and an event start tag is inserted at the beginning of each sliding time window. A user anchor tag is inserted at the first appearance of the analyzed user to obtain the metaverse interaction long sequence.

[0079] Step S212: Construct multi-channel embeddings using meta-universe interaction long sequences to obtain input representation vectors. Multi-channel embeddings include word action embedding, absolute position embedding, time interval embedding, role relationship embedding, and social structure embedding.

[0080] Word-behavior embedding is used to construct word embedding matrices and behavior tag embedding matrices for text words and behavior tags, respectively.

[0081] Absolute position embedding is used for position index encoding based on markers in a long sequence of metacosmic interactions;

[0082] The time interval embedding is used to segment and encode based on the time difference between the behavior time corresponding to the long sequence of metaverse interactions and the starting time of the sliding window;

[0083] Role relationship embedding is used to distinguish different conversational participants such as target users, their own digital avatars, user friends, friends in the same community, and strangers;

[0084] Social structure embedding is used to map friend edge weights and community edge weights into low-dimensional vectors based on the friend attention network and community participation network;

[0085] Step S213: Input the input representation vector into the precoding layer to obtain the initial hidden representation. Calculate the global importance score for each tag based on the initial hidden representation. The expression for calculating the global importance score is:

[0086] ;

[0087] in, For the Sigmoid function, This is the initial hidden representation. and For learnable parameters, A global importance score is set with a preset score threshold. When the global importance score is greater than the score threshold, the current corresponding label is set as the global attention label, and the remaining labels are treated with local sliding window attention to generate a sparse attention structure.

[0088] Step S214: Longformer encoding and gated self-attention pooling are used to perform deep representation processing on the initial hidden representation to obtain the text feature vector.

[0089] User behavior event data within the time window is concatenated in ascending order of timestamps to generate a basic text sequence. A behavior encoding dictionary is defined to map non-text behaviors to special markers, such as [GIFT] for giving a gift and [FOCUS] for staring for a long time. Based on the time point of the behavior, the behavior marker is inserted into the corresponding position in the text sequence. Among the special markers, [MSG] is inserted before each original text message to identify the message boundary, [EVENTSTART] is inserted at the starting position of the sliding window, and [ANCHORUSER] is inserted at the first appearance position of the target user. The sequence format is unified to obtain a metaverse interaction long sequence, ensuring that the length does not exceed the processing limit of the Longformer model.

[0090] Construct multi-channel embeddings, with the first channel being word-behavior embeddings: using a pre-trained BERT vocabulary, initialize the word embedding matrix and randomly initialize the behavior embedding matrix, and select the corresponding embedding matrix for lookup based on the tag type;

[0091] The second channel is absolute position embedding: it uses learnable position encoding to assign a unique position vector to each position in the sequence, with position indices ranging from 0 to sequence length - 1;

[0092] The third channel is time interval embedding: the time difference between the timestamp of each tag and the window start time is calculated and processed using piecewise linear encoding;

[0093] The fourth channel is role relationship embedding: Define a set of roles {target user, self clone, close friend, ordinary friend, same community, stranger}, assign a learnable embedding vector to each role, and define role labels based on the social relationships of the user's behavior initiator;

[0094] The fifth channel is social outcome embedding: embedding through friend attention networks and community engagement networks;

[0095] By concatenating the five channels, the final input representation vector is:

[0096] ;

[0097] in, The input representation vector, Embedded as a word behavior, For absolute position embedding, Embedded for time intervals, Embedding for character relationships, To embed social results, this embodiment unifies multi-source information such as semantics, temporal sequence, social interaction, and roles into the same vector space, preserves fine-grained differences, encodes different types of information through independent channels, avoids feature confusion, and provides a rich feature foundation for subsequent sparse attention calculation.

[0098] The input vector sequence X = [x1, x2, ..., xn] is given, with a preset score threshold of 0.3. Vectors with a global importance score greater than 0.3 are labeled with global attention, while those with a global importance score less than or equal to 0.3 are labeled with local attention. For example, assuming a sequence length of 1000, the importance score of each label is calculated, and high-scoring labels are obtained as follows:

[0099] [ANCHORUSER](0.92),[SHAREDOC](0.85),[ASSIGN_TASK](0.78);

[0100] Medium score marker: Main discussion points (0.4-0.6);

[0101] Low score markers: modal particles, filler words (0.1-0.2);

[0102] If the score threshold is set to 0.3, the first three tags are selected as global attention tags, and the remaining tags can only be focused on within 64 positions before and after them;

[0103] After 12 layers of Longformer encoding, each token in the sequence obtains a deep context identifier, enabling user-tailored feature extraction by embedding from social networks into a personalized attention mechanism.

[0104] Step S3 includes the following sub-steps:

[0105] Step S31: Standardize and encode the text classification probability and behavioral context to obtain the text classification vector and behavioral context vector;

[0106] Step S32: Using a hybrid kernel embedding learning model, the text classification vector and the behavior context vector are fused and analyzed to output user behavior mapping data. The user behavior mapping data includes user behavior parameters and context labels. The user behavior parameters include social initiative parameters, emotional expression intensity parameters, topic preference parameters, response speed parameters, response length parameters, emotional matching degree parameters, and task engagement degree parameters.

[0107] This embodiment is the core innovation of the present invention. Its purpose is to deeply integrate the text classification probability output in step S2 with the behavioral context (environment and situation in which the interaction occurs) collected in step S1 to generate a set of user behavior mapping data with quantifiable structure. This data directly defines the specific behavioral performance parameters of the digital clone in various situations and is the fundamental basis for realizing personalized behavior generation.

[0108] The processing logic for fusing and analyzing text classification vectors and behavioral context vectors using a hybrid kernel embedding learning model is as follows:

[0109] The text classification vectors are mapped to a high-dimensional semantic space using a deep variational Gaussian kernel to obtain semantic distribution features. Behavioral context vectors are constructed using a spectral graph convolutional network to create a structural representation, which is then mapped using a Gaussian process to obtain behavioral structural features. Finally, a hybrid semantic kernel embedding is used to fuse the semantic distribution features and behavioral structural features, generating hybrid semantic features. The calculation expression for the hybrid semantic kernel embedding is as follows:

[0110] ;

[0111] in, It is a hybrid semantic feature. Semantic distribution features As behavioral structural features, For fusion weights;

[0112] The hybrid semantic features are input into the behavior parameter decoder, which outputs user behavior parameters. The user behavior parameters are classified according to the triple self-supervised scenario optimization, and the corresponding scenario labels are obtained. The user behavior parameters and scenario labels are saved as user behavior mapping data.

[0113] The essence of obtaining the probability distribution is that the text classification vector itself is a discrete distribution. By using a deep variational Gaussian kernel, the text classification vector is mapped to a continuous latent space, and the similarity between the text classification probability and a series of prototype distributions is calculated. This allows for a more nuanced characterization of the user's intent patterns. The variational method allows us to model text classification vectors that may contain noise or uncertainty (e.g., intent ambiguity when the probability value is close to 0.5), making the features more robust. For example, when the text classification vector = [-1.2, 2.1, 0.3] (high sentiment), after deep network mapping and variational sampling, a semantic distribution feature (a 64-dimensional vector) is obtained. If there are 5 prototype vectors representing high utilitarianism, high sentiment, balance, utilitarianism with secondary sentiment, and sentiment with secondary utilitarianism, the Euclidean distance between the semantic distribution feature and these 5 prototypes is calculated and converted into a similarity using a Gaussian kernel. The final result is... The semantic distribution features are [0.1, 0.8, 0.05, 0.03, 0.7]. This indicates that the current text intent is most similar to a highly emotional prototype, and also relatively similar to a text primarily driven by emotion with secondary utilitarianism.

[0114] The elements in a behavioral context are not isolated but interconnected. Spectral graph convolution can explicitly model the relationships between these elements, extracting structured features, i.e., behavioral structure features. These represent the degree of matching between the current context and some typical scene structures, such as private leisure and public formality. When the graph structure corresponding to the behavioral context vector is: the target user is connected to four nodes—late night, private beach, close friend A, and friend B—after passing through two layers of GCN, the feature encoding of the central node encodes information about the graph structure. Given four basic structural patterns: private leisure, public work, large-scale social interaction, and solitary exploration, the similarity between the features of the central node and these four patterns is calculated to obtain... Behavioral structure features = [0.9, 0.1, 0.2, 0.05]. This indicates that the current context highly matches the private leisure structure;

[0115] Adopting and solving its criteria ,in =1, If all values ​​are greater than 0, construct a target similarity matrix and obtain the optimal weights through iterative optimization using projection gradient ascent. =0.6, and the final output user behavior mapping data consists of context labels and user behavior parameters;

[0116] In this embodiment, dual-path fusion takes into account both the user's individual intent and the specific context, resulting in digital clone behavior that truly matches the user's habits in specific situations. The hybrid kernel model can handle unseen combinations of intent scenarios and provides reasonable behavioral parameters through generalization, enabling the clone to behave naturally in new scenarios. Through triple self-supervised scenario optimization, behavioral scenarios are automatically discovered and labeled, reducing the cost of manual labeling. Furthermore, scenario labels help organize behavioral mapping data, facilitating subsequent retrieval and scenario matching. Through clustering and contrastive learning, samples under the same scenario are clustered in the feature space, while different scenarios are separated, thus giving the scenario labels excellent discriminativeness and representativeness.

[0117] The logic for classifying user behavior parameters based on triple self-supervised scenario optimization and obtaining corresponding scenario labels is as follows:

[0118] Calculate the soft assignment probability from the mixed semantic features to the corresponding context labels. The expression for calculating the soft assignment probability is as follows:

[0119]

[0120] in, Let i be the hybrid semantic kernel embedding vector of the i-th sample. Let the vector centered at the c-th row represent a typical behavioral pattern. Degrees of freedom allocated to behavior The soft-assignment probability of the mixed semantic features of the i-th sample corresponding to the c-th row is given.

[0121] The target distribution is calculated from the soft allocation probability. The expression for the target distribution is:

[0122]

[0123] in, For the target distribution, The soft-assigned probabilities of all samples i are summed; the output scenario category labels are optimized by using an adversarial loss function in conjunction with gradient descent, and user behavior parameters are mapped one-to-one with the corresponding scenario category labels to generate user behavior mapping data.

[0124] By minimizing the difference between the soft assignment probability and the target distribution, the sample embeddings within the same cluster are made more compact, while the samples used in different clusters are more separated.

[0125] Step S4 includes the following sub-steps:

[0126] Step S41: Obtain the current environment data of the user when offline from the metaverse environment interface, preprocess the current environment data, and generate the current environment feature vector;

[0127] Step S42: Using the scenario category matching algorithm, calculate the cosine similarity between the current environment feature vector and the center of each row in the user behavior mapping data, and select the behavior center with the highest cosine similarity as the current scenario category label.

[0128] Step S43: Based on the current scenario category label, retrieve all user behavior parameters belonging to the scenario category label from the user behavior mapping data, and perform a weighted average of the user behavior parameters to obtain the user behavior parameters in the current environment.

[0129] In this embodiment, by acquiring the user's offline environment data, the digital clone can autonomously perceive the environment, identify the scenario, and generate personalized behaviors that conform to the user's habits. Through an efficient scenario matching and parameter fusion mechanism, it is ensured that the output vector and the offline learning behavior center are in the same vector space, realizing a smooth transformation from static behavioral knowledge to dynamic behavioral performance.

[0130] The processing logic of the scenario category matching algorithm is as follows:

[0131] The current environment data is preprocessed to generate the cosine similarity between the current environment feature vector and the feature vector corresponding to each scenario category label. The scenario category with the highest cosine similarity is selected as the current scenario category label.

[0132] This embodiment is a fast scene recognition system based on a vector space model. The purpose is to transform the environmental perception problem into a nearest neighbor search problem in a high-dimensional vector space. The environmental feature vector must be located in the same embedding space as the behavior center vector. The meaning of the scene label remains consistent in offline learning and online matching. After matching a scene, it can be backtracked to the corresponding behavior parameter set. This embodiment uses cosine similarity measurement to transform the complex pattern recognition problem into a similarity calculation problem in a vector space, thereby improving the naturalness of the digital clone's behavior and user satisfaction.

[0133] Context category tags include:

[0134] Collaborative work, social entertainment, exploratory learning, and leisure observation.

[0135] By pre-setting four basic scenario tags, a behavioral pattern classification system covering the core activity types of the metaverse was constructed, realizing the manageability and interpretability of complex user behaviors. This helps to simplify the computational complexity of behavior matching, ensure real-time response performance, and, through a semantically clear tag system, make the behavioral decisions of digital avatars transparent and easy to understand for human users, significantly improving the trust and naturalness of human-computer collaboration.

[0136] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention. All data acquisition actions in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located and with the authorization granted by the owner of the corresponding device.

Claims

1. A method for generating user digital avatar behavior patterns based on the metaverse, characterized in that, Includes the following steps: Step S1: Process the collected user social data, metaverse behavior data, and behavior context to obtain user behavior event data. The user social data includes text data, image data, and interactive behavior data. The metaverse behavior data includes micro-expressions, action data, and social behavior data. The behavior context includes timestamps, locations, scenes, activity types, and interacting users. Step S2: Process the user behavior event data using a multi-channel text feature model to obtain text feature vectors, and obtain text classification probabilities through multi-channel task processing. Step S3: Construct a hybrid kernel embedding learning model to map text classification probabilities and behavioral context to obtain user behavior mapping data; Step S4: Obtain user behavior parameters by analyzing user behavior mapping data and the current metaverse environment; Step S2 includes the following sub-steps: Step S21: Use a multi-channel text feature model to perform self-attention pooling on the text data in the user behavior event data to obtain text feature vectors; Step S22: Map the text feature vectors to a shared low-dimensional semantic space to obtain low-dimensional shared semantic vectors; Step S23: The output low-dimensional shared semantic vector is classified using a multi-head task parallel channel to obtain the text classification probability, which includes the utilitarian imbalance probability, the emotional imbalance probability, and the normal interaction probability. Step S3 includes the following sub-steps: Step S31: Standardize and encode the text classification probability and behavioral context to obtain the text classification vector and behavioral context vector; Step S32: Using a hybrid kernel embedding learning model, the text classification vector and the behavior context vector are fused and analyzed to output user behavior mapping data. The user behavior mapping data includes user behavior parameters and context labels. The user behavior parameters include social initiative parameters, emotional expression intensity parameters, topic preference parameters, response speed parameters, response length parameters, emotional matching degree parameters, and task engagement degree parameters. The processing logic for fusing and analyzing text classification vectors and behavioral context vectors using a hybrid kernel embedding learning model is as follows: The text classification vector is mapped to a high-dimensional semantic space using a deep variational Gaussian kernel to obtain semantic distribution features. The behavioral context vector is constructed using a spectral graph convolutional network to create a structural representation, which is then mapped using a Gaussian process to obtain behavioral structural features. Finally, a hybrid semantic kernel embedding is used to fuse the semantic distribution features and behavioral structural features to generate hybrid semantic features. The calculation expression for the hybrid semantic kernel embedding is as follows: ； in, It is a hybrid semantic feature. Semantic distribution features As behavioral structural features, For fusion weights; The hybrid semantic features are input into the behavior parameter decoder, which outputs user behavior parameters. The user behavior parameters are classified according to the triple self-supervised scenario optimization, and the corresponding scenario labels are obtained. The user behavior parameters and scenario labels are saved as user behavior mapping data.

2. The method for generating user digital avatar behavior patterns based on the metaverse as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: Obtain user social data, metaverse behavior data, and behavior context through the data collection SDK built into the metaverse client; Step S12: Construct a social network structure using user social data. The social network structure includes a friend attention network and a community participation network. The friend attention network uses the user as the main node and each friend user as the edge node. The friend edge weight is used to connect the user with each friend user. The friend edge weight is the proportion of the user's chat time with a single friend user to the user's total chat time within a unit period. The community participation network uses the user as the main node and each community as the edge node. The community edge weight is used to connect the user with each community. The community edge weight is the proportion of the user's chat time in a single community to the user's chat time in all communities within a unit period. Step S13: Match user social data, metaverse behavior data, and behavior context one-to-one according to the same timestamp to obtain metaverse interaction data. Set a sliding time window and divide the metaverse interaction data into user behavior event data according to the sliding time window.

3. The method for generating user digital clone behavior patterns based on the metaverse as described in claim 2, characterized in that, The calculation expression for the low-dimensional shared semantic vector is as follows: ； in, For activation function, For layer normalization, For low-dimensional shared semantic vectors, and For learning parameters, This is the text feature vector.

4. The method for generating user digital clone behavior patterns based on the metaverse as described in claim 3, characterized in that, Step S21 further includes the following sub-steps: Step S211: The text data that has been divided in the behavioral event data is concatenated in chronological order to construct a long text sequence. At the same time, the interactive behavioral data and metaverse behavioral data that occur at the corresponding time points are converted into predefined behavioral tags and inserted into the time positions corresponding to the long text sequence to generate a long sequence. The long sequence includes text data and metaverse behavioral data. A session unit tag is inserted before each text message in the long text sequence. A sliding time window is set. An event start tag is inserted at the beginning of each sliding time window. A user anchor tag is inserted at the first appearance position of the analyzed user to obtain a metaverse interaction long sequence. Step S212: Construct a multi-channel embedding using the meta-universe interaction long sequence to obtain the input representation vector. The multi-channel embedding includes word action embedding, absolute position embedding, time interval embedding, role relationship embedding, and social structure embedding. The word-behavior embedding is used to establish word embedding matrices and behavior tag embedding matrices for text words and behavior tags, respectively; The absolute position embedding is used to encode the position index based on the markers in the metaverse interaction long sequence; The time interval embedding is used to perform segmented encoding based on the time difference between the behavior time corresponding to the long sequence of metaverse interactions and the starting time of the sliding window; The role relationship embedding is used to distinguish different conversational participation roles of target users, their own digital clones, user friends, friends in the same community, and strangers. The social structure embedding is used to map friend edge weights and community edge weights into low-dimensional vectors based on the friend attention network and community participation network. Step S213: Input the input representation vector into the precoding layer to obtain the initial hidden representation. Calculate the global importance score for each tag based on the initial hidden representation. The expression for calculating the global importance score is: ； in, For the Sigmoid function, This is the initial hidden representation. and For learnable parameters, A global importance score is set with a preset score threshold. When the global importance score is greater than the score threshold, the current corresponding label is set as the global attention label, and the remaining labels are treated with local sliding window attention to generate a sparse attention structure. Step S214: Longformer encoding and gated self-attention pooling are used to perform deep representation processing on the initial hidden representation to obtain the text feature vector.

5. The method for generating user digital clone behavior patterns based on the metaverse as described in claim 4, characterized in that, The logic for classifying user behavior parameters based on triple self-supervised scenario optimization and obtaining corresponding scenario labels is as follows: Calculate the soft assignment probability of the mixed semantic features to the corresponding context labels. The expression for calculating the soft assignment probability is as follows: ； in, Let i be the hybrid semantic kernel embedding vector of the i-th sample. Let the vector centered at the c-th row represent a typical behavioral pattern. Degrees of freedom allocated to behavior The soft-assignment probability of the mixed semantic features of the i-th sample corresponding to the c-th row is given. The target distribution is calculated for the soft allocation probability, and the expression for the target distribution is as follows: ； in, For the target distribution, The soft-assignment probabilities of all samples i are summed; the output scenario category labels are optimized by using an adversarial loss function and gradient descent, and user behavior parameters are mapped one-to-one with the corresponding scenario category labels and linked to all behavior centers. Save as user behavior mapping data.

6. The method for generating user digital clone behavior patterns based on the metaverse as described in claim 5, characterized in that, Step S4 includes the following sub-steps: Step S41: Obtain the current environment data of the user when offline from the metaverse environment interface, preprocess the current environment data, and generate the current environment feature vector; Step S42: Using the scenario category matching algorithm, calculate the cosine similarity between the current environment feature vector and the center of each row in the user behavior mapping data, and select the behavior center with the highest cosine similarity as the current scenario category label. Step S43: Based on the current scenario category label, retrieve all user behavior parameters belonging to the scenario category label from the user behavior mapping data, and perform a weighted average of the user behavior parameters to obtain the user behavior parameters in the current environment.

7. The method for generating user digital clone behavior patterns based on the metaverse as described in claim 6, characterized in that, The processing logic of the scenario category matching algorithm is as follows: The current environment data is preprocessed to generate the cosine similarity between the current environment feature vector and the feature vector corresponding to each scenario category label. The scenario category with the highest cosine similarity is selected as the current scenario category label.

8. The method for generating user digital clone behavior patterns based on the metaverse as described in claim 7, characterized in that, The scenario category labels include: Collaborative work, social entertainment, exploratory learning, and leisure observation.

Images