User portrait construction method under AI platform

Abstract

The invention provides a user portrait construction method under an AI platform, and belongs to the technical field of user portray.The method comprises the steps that microscopic interaction data are obtained through a user behavior collector, and a behavior state vector set is constructed according to a time sequence; a probability model and an interest energy function are established based on a Maxwell-Boltzmann distribution principle to automatically identify and filter noise data, and sparse matrix decomposition is adopted to project filtered clean data to a low-dimensional semantic feature space. A semantic enhancement model is utilized to perform clustering analysis on high-quality behavior data to identify a real user interest mode, a user individual graph structure based on semantic association is constructed through a knowledge graph technology to accurately express user preferences, and MCP protocol packaging is adopted to realize fusion processing of clean microscopic behaviors and accurate macroscopic features. The technical problem that the accuracy of the formed user portrait is not enough due to the fact that the user microcosmic behavior data is easily interfered by noise is solved.

Description

Technical Field

[0001] This invention belongs to the field of user profiling technology, and more specifically, relates to a method for constructing user profiles on an AI platform. Background Technology

[0002] User profiling is a core technology for AI platforms to achieve accurate recommendations and personalized services. Traditional methods mainly employ collaborative filtering algorithms to discover user preference patterns by analyzing user-item rating matrices, or deep learning-based neural network models to learn feature representations from users' historical behavior sequences. These technologies are widely used in e-commerce platform product recommendations, video website content distribution, social media advertising, and personalized learning in online education, predicting users' future interests by mining regular features in user behavior data. However, in practical applications, user micro-behavioral data inevitably contains a large amount of noise, including accidental clicks, unintentional repeated refreshes, malicious user boosting behavior, bot programs' simulated operations, and repeated requests due to network latency. This noisy data is unrelated to users' true interests but is treated equally by traditional algorithms, severely interfering with the accuracy of user profiling. Traditional collaborative filtering methods directly perform similarity calculations and matrix factorization based on raw behavioral data, failing to effectively distinguish between genuine interest behaviors and noise interference. While deep learning methods have certain feature learning capabilities, they lack targeted noise identification and filtering mechanisms, often treating noisy data as effective features, leading to overfitting and decreased generalization ability in the model. In current user modeling practices, due to the lack of effective data quality control and noise reduction techniques, user profiles constructed using existing methods often deviate from users' true interests and exhibit significant inaccuracies when faced with complex data environments. In other words, existing technologies suffer from the technical problem that user micro-behavioral data is easily interfered with by noise, leading to inaccurate user profiles. Summary of the Invention

[0003] In view of this, the present invention provides a user profile construction method under an AI platform, which can solve the technical problem in the prior art that user micro-behavioral data is easily interfered with by noise, resulting in insufficient accuracy of the user profile.

[0004] This invention is implemented as follows: It provides a method for constructing user profiles on an AI platform, comprising: acquiring real-time micro-interaction data of users on the platform through a user behavior collector; arranging the micro-interaction data according to a time series to form a set of user behavior micro-state vectors; constructing a user behavior state probability model based on the Maxwell-Boltzmann distribution principle in statistical mechanics; mapping the set of user behavior micro-state vectors to a macro-statistical feature space; establishing a user interest energy function by calculating the distribution entropy value of each behavior dimension; using a sparse matrix to represent the relationship between users and content entities; and using a matrix factorization algorithm to transform high-dimensional user behavior... The process involves projecting the spatial model onto a low-dimensional semantic feature space; using a user profile semantic enhancement model to perform cluster analysis on the sparse matrix to identify micro-clustering units of users with similar behavioral patterns; extracting entity relationship pairs from user behavior content based on knowledge graph triple extraction technology to construct individual user semantic graph structures; encapsulating user profile modeling services and knowledge graph construction services through the MCP protocol to achieve the fusion processing of user behavior micro-state vector sets and macro-profile features; and establishing a dynamic update mechanism for user profiles to adjust user interest energy function parameters and individual user semantic graph structure weights in real time based on newly added behavioral data to complete user profile construction.

[0005] The micro-interaction data includes click coordinates, dwell time, swipe trajectory, input content, and voice interaction frequency. The steps to establish the user behavior micro-state vector set are as follows: normalize the micro-interaction data to eliminate dimensional differences, classify and label the micro-interaction data according to behavior type to form behavior type identifiers, combine multiple behavior type identifiers within the same time window into a state vector, and arrange the state vectors in chronological order to form the user behavior micro-state vector set.

[0006] The user interest energy function is used to quantify the attractiveness of different content types to users. The specific steps for constructing the user interest energy function are to calculate the probability distribution of the occurrence frequency of each behavioral dimension, calculate the energy value of each behavioral state based on the Maxwell-Boltzmann distribution formula, weight and fuse the energy value of the behavioral state with the semantic features of the content, and adjust the time weight of historical behaviors through an exponential decay function to obtain the final user interest energy function.

[0007] In this sparse matrix, rows represent macro-level user group classifications, columns represent micro-level content entities, and matrix element values ​​represent association strength weights. The optimization steps of sparse matrix decomposition specifically involve using singular value decomposition (SVD) to reduce the dimensionality of the user-content association matrix, iteratively optimizing the loss function of matrix decomposition using alternating least squares, setting regularization parameters to prevent overfitting and maintain sparsity, and using the decomposed user feature matrix and content feature matrix as the basic representation of the low-dimensional semantic feature space.

[0008] Each user micro-cluster unit corresponds to a combination of interest preferences. The steps for identifying user micro-cluster units are as follows: calculate the cosine similarity distance matrix based on the user feature vector, use the density peak clustering algorithm to identify user groups with similar behavioral patterns, calculate the center vector and variance parameter for each cluster unit, and establish the mapping relationship between cluster units and interest tags to form a micro-user profile template.

[0009] Among them, the attention weight parameters of the user profile semantic enhancement model are dynamically adjusted according to three parameters: user historical interaction frequency, content semantic similarity, and time decay factor. The user profile semantic enhancement model is a multi-head attention network based on the Transformer architecture, which includes four main components: user behavior encoding layer, semantic feature extraction layer, attention fusion layer and profile generation layer. The attention fusion layer adopts an adaptive weight mechanism to dynamically adjust the importance of each feature dimension according to the user interaction pattern.

[0010] The user-individual semantic graph structure includes four types: user nodes, content nodes, topic nodes, and tag nodes. The specific steps for constructing the user-individual semantic graph structure are as follows: extract named entities and relation pairs from user interaction content, calculate the semantic similarity between entities based on a pre-trained language model (such as BERT), construct a user-centric multi-hop semantic relation graph, assign semantic type labels and weight attributes to each node in the graph, and obtain the final user-individual semantic graph structure through graph neural network node representation learning.

[0011] Specifically, the process involves transforming the set of micro-state vectors representing user behavior into a macro-knowledge representation through semantic relationship links, jointly optimizing the semantic graph structure of individual users with the macro-statistical model of the global user group, and defining standardized user profile data interface specifications through the MCP protocol encapsulation process. This includes encapsulating the set of micro-state vectors representing user behavior and macro-statistical features into a unified message format, establishing an asynchronous communication mechanism between services to support real-time data exchange, and allocating computing resources through a load balancing algorithm to ensure system stability.

[0012] The user profile dynamic update mechanism maintains the timeliness and accuracy of the profile model through a time window sliding strategy. The specific steps of the user profile dynamic update mechanism are to set incremental update triggers at fixed time intervals, calculate the impact weight of new behavioral data on the existing profile model, update the user interest energy function parameters using an online learning algorithm, update the node and edge weights of the individual user semantic graph structure through graph structure modification operations, and record the profile change history to support model rollback and version management.

[0013] The steps for establishing the training dataset for the user profile semantic enhancement model include collecting large-scale user behavior log data covering different fields and time periods, cleaning and standardizing the behavior data to remove outliers and noise, constructing user behavior sequence samples and labeling them with corresponding profile tags, expanding the diversity of training samples through data augmentation techniques, and dividing the training set, validation set, and test set to ensure the rationality of data distribution.

[0014] The user profile semantic enhancement model takes a set of user behavior micro-state vectors and content semantic embedding vectors as inputs and outputs a multi-dimensional user profile feature representation vector. The training steps of the user profile semantic enhancement model include using the Adam optimization algorithm to update model parameters through gradient descent, setting a learning rate decay strategy to prevent oscillations during training, using a weighted combination of cross-entropy loss function and mean squared error loss function as the optimization objective, and preventing overfitting through early stopping mechanism and model checkpoint saving.

[0015] Among them, the frequency of user historical interaction refers to the cumulative number of times a user interacts with similar content in the past 30 days, which is used to measure the user's continued attention to the content type. The semantic similarity of content refers to the degree of semantic association between different content entities calculated by cosine similarity, with a value range of 0 to 1.

[0016] Among them, the time decay factor refers to the historical behavior weight decay coefficient calculated based on the exponential function, which is used to reduce the impact of long-term behavior on the current profile. The low-dimensional semantic feature space refers to the compact representation space obtained by reducing the dimensionality of the original high-dimensional feature vector through matrix factorization. The dimension is usually one-tenth of the original dimension.

[0017] Among them, the micro user profile template refers to a standardized user behavior pattern description framework obtained based on cluster analysis, which includes interest weight distribution and behavior sequence features. By analogy of user micro behavior to the random movement of gas molecules, a mathematical mapping relationship from individual behavior to group statistical characteristics is established.

[0018] In this approach, each user's click, browse, and stay are treated as a molecular motion state, with each behavior state corresponding to an energy level. High-frequency behaviors correspond to low-energy states, while low-frequency behaviors correspond to high-energy states. By statistically analyzing the probability distribution of these micro-states, macro-level user interests and preferences are derived, thus realizing the transformation from massive discrete behavioral data to continuous probability distributions.

[0019] At the micro level, the behavior sequence of an individual user exhibits randomness and uncertainty, while at the macro level, it demonstrates stable statistical regularity through the law of large numbers. The Maxwell-Boltzmann distribution can naturally describe the equilibrium state and fluctuations of the system, corresponding to the stability and variability of user interests. When a user pays long-term attention to a certain type of content, the system tends to be in equilibrium, while fluctuations occur when user interests shift.

[0020] This invention constructs a probabilistic model of user behavior states by introducing the Maxwell-Boltzmann distribution principle from statistical mechanics. It analogizes user micro-interactions to the random motion of gas molecules, and automatically identifies and filters noisy data that deviates from normal distribution patterns by statistically analyzing the probability distributions of a large number of behavior states. This achieves a transformation from noise-sensitive direct data processing to probabilistic statistical modeling with inherent noise filtering capabilities. The invention uses a user interest energy function to quantify the energy levels of different behavior states. Real user interest behaviors follow stable energy distribution patterns, while noisy behaviors, due to their randomness and anomalies, exhibit extreme energy states. By setting reasonable energy thresholds, these abnormal data points can be effectively distinguished and filtered, significantly improving the quality and reliability of behavior data. This invention combines sparse matrix factorization and knowledge graph technology to perform deep understanding and feature enhancement on filtered clean data through semantic association. The constructed user individual semantic graph structure can accurately reflect the user's true interests and preferences, avoiding interference and pollution of profile features by noisy data. At the same time, the dynamic update mechanism continuously optimizes data quality, maintaining a high degree of consistency between user profile and true interests. In summary, this invention solves the technical problem mentioned in the background technology that user micro-behavioral data is easily interfered with by noise, resulting in insufficient accuracy of the user profile. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the user profile semantic enhancement model structure involved in the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] like Figure 1 The diagram shown is a flowchart of a user profile construction method under an AI platform provided by the present invention. This method includes the following steps:

[0025] S01. The micro-interaction data of users on the platform is acquired in real time through the user behavior collector, including click coordinates, dwell time, swipe trajectory, input content, and voice interaction frequency. The micro-interaction data is arranged in time series to form a set of user behavior micro-state vectors.

[0026] S02. Based on the Maxwell-Boltzmann distribution principle in statistical mechanics, a user behavior state probability model is constructed, and the set of user behavior micro-state vectors is mapped to a macro-statistical feature space. By calculating the distribution entropy value of each behavior dimension, a user interest energy function is established. The user interest energy function is used to quantify the attraction intensity of different content types to users.

[0027] S03. A sparse matrix is ​​used to represent the relationship between users and content entities. The rows of the sparse matrix represent macro-level user group classifications, the columns represent micro-level content entities, and the matrix element values ​​represent the association strength weights. The high-dimensional user behavior space is projected to the low-dimensional semantic feature space through a matrix factorization algorithm.

[0028] S04. Use the user profile semantic enhancement model to perform cluster analysis on the sparse matrix to identify user micro-clustering units with similar behavioral patterns. Each user micro-clustering unit corresponds to an interest preference combination. The attention weight parameter of the user profile semantic enhancement model is dynamically adjusted according to three parameters: user historical interaction frequency, content semantic similarity, and time decay factor.

[0029] S05. Based on the knowledge graph triplet extraction technology, entity relationship pairs are extracted from user behavior content to construct a user individual semantic graph structure. The user micro-behavior is transformed into macro-knowledge representation through semantic relationship links. The user individual semantic graph structure includes four types: user nodes, content nodes, topic nodes, and tag nodes.

[0030] S06. User profile modeling service and knowledge graph construction service are encapsulated through the MCP protocol to realize the fusion processing of the user behavior micro-state vector set and macro profile features, and to jointly optimize the individual semantic graph structure of individual users and the macro statistical model of the global user group.

[0031] S07. Establish a dynamic update mechanism for user profiles, adjust the parameters of the user interest energy function and the weights of the individual user semantic graph structure in real time based on newly added behavioral data, and maintain the timeliness and accuracy of the profile model through a time window sliding strategy.

[0032] The steps for establishing the user behavior micro-state vector set specifically include: normalizing the micro-interaction data to eliminate dimensional differences; classifying and labeling the micro-interaction data according to behavior type to form behavior type identifiers; combining multiple behavior type identifiers within the same time window into a state vector; and arranging the state vectors in chronological order to form the user behavior micro-state vector set.

[0033] The steps for constructing the user interest energy function specifically include: calculating the probability distribution of the occurrence frequency of each behavioral dimension; calculating the energy value of each behavioral state based on the Maxwell-Boltzmann distribution formula; weighting and fusing the energy value of the behavioral state with the semantic features of the content; and adjusting the time weight of historical behaviors through an exponential decay function to obtain the final user interest energy function.

[0034] The optimization steps of the sparse matrix decomposition specifically include: using the singular value decomposition algorithm to reduce the dimensionality of the user content association matrix; iteratively optimizing the loss function of the matrix decomposition using the alternating least squares method; setting a regularization parameter to prevent overfitting and maintain sparsity; and using the decomposed user feature matrix and content feature matrix as the basic representation of the low-dimensional semantic feature space.

[0035] The steps for identifying user micro-clustering units specifically include: calculating a cosine similarity distance matrix based on user feature vectors; identifying user groups with similar behavioral patterns using density peak clustering algorithm; calculating the center vector and variance parameter for each clustering unit; and establishing a mapping relationship between clustering units and interest tags to form a micro-user profile template.

[0036] The steps for constructing the user individual semantic graph structure specifically include: extracting named entities and relation pairs from user interaction content; calculating the semantic similarity between entities based on a pre-trained language model; constructing a user-centric multi-hop semantic relation graph; assigning semantic type labels and weight attributes to each node in the graph; and obtaining the final user individual semantic graph structure through node representation learning using a graph neural network.

[0037] The specific steps of the MCP protocol encapsulation process include: defining a standardized user profile data interface specification; encapsulating the user behavior micro-state vector set and macro-statistical features into a unified message format; establishing an asynchronous communication mechanism between services to support real-time data exchange; and allocating computing resources through a load balancing algorithm to ensure system stability.

[0038] The implementation steps of the dynamic update mechanism for user profiles specifically include: setting incremental update triggers at fixed time intervals; calculating the impact weights of new behavioral data on the existing profile model; updating the parameters of the user interest energy function using an online learning algorithm; updating the node and edge weights of the individual user semantic graph structure through graph structure modification operations; and recording the profile change history to support model rollback and version management.

[0039] The specific structure of the user profile semantic enhancement model is a multi-head attention network based on the Transformer architecture, comprising four main components: a user behavior encoding layer, a semantic feature extraction layer, an attention fusion layer, and a profile generation layer. The attention fusion layer employs an adaptive weighting mechanism to dynamically adjust the importance of each feature dimension according to user interaction patterns. The model input consists of the set of user behavior micro-state vectors and content semantic embedding vectors, and the output is a multi-dimensional user profile feature representation vector. The training dataset establishment steps for the user profile semantic enhancement model specifically include collecting large-scale user behavior log data covering different domains and time periods, cleaning and standardizing the behavior data to remove outliers and noise, constructing user behavior sequence samples and labeling them with corresponding profile tags, expanding the diversity of training samples through data augmentation techniques, and dividing the training set, validation set, and test set to ensure the rationality of data distribution. The training steps for the user profile semantic enhancement model specifically include using the Adam optimization algorithm for gradient descent to update model parameters, setting a learning rate decay strategy to prevent oscillations during training, using a weighted combination of cross-entropy loss function and mean squared error loss function as the optimization objective, preventing overfitting through early stopping mechanisms and model checkpoint saving, evaluating model performance on the validation set, and adjusting hyperparameters until convergence.

[0040] The user's historical interaction frequency refers to the cumulative number of interactions a user has had with similar content over the past 30 days, used to measure the user's sustained attention to content types. The content semantic similarity refers to the degree of semantic association between different content entities calculated using cosine similarity, with a value ranging from 0 to 1. The time decay factor refers to the historical behavior weight decay coefficient calculated based on an exponential function, used to reduce the impact of long-standing behaviors on the current profile. The low-dimensional semantic feature space refers to the compact representation space obtained by reducing the dimensionality of the original high-dimensional feature vector through matrix factorization; its dimension is typically one-tenth of the original dimension. The micro-user profile template refers to a standardized user behavior pattern description framework obtained based on clustering analysis, including interest weight distribution and behavioral sequence features.

[0041] This invention employs the Maxwell-Boltzmann distribution principle from statistical mechanics to address the micro-to-macro correlation problem in user profiling modeling. By analogying user micro-behaviors to the random motion of gas molecules, a mathematical mapping relationship is established from individual behavior to group statistical characteristics. Physical principles have been widely applied in modeling socio-economic phenomena; Brownian motion models are used for stock price prediction, the concept of thermodynamic entropy is introduced into information theory, the Ising model is applied to social network analysis, and gravity models are used for population migration prediction. These successful cases demonstrate the effectiveness of physical laws in complex system modeling. These physical laws play a crucial role in data modeling, treating each user's click, browse, or dwell time as a molecular motion state. Each behavioral state corresponds to an energy level; high-frequency behaviors correspond to low-energy states, and low-frequency behaviors correspond to high-energy states. By statistically analyzing the probability distribution of these micro-states, macro-level user interests and preferences are derived, achieving the transformation from massive discrete behavioral data to continuous probability distributions. At the micro level, the behavioral sequence of a single user exhibits randomness and uncertainty, but at the macro level, it displays stable statistical regularities through the law of large numbers, just as the trajectories of individual molecules in a gas are irregular, but the overall temperature and pressure remain stable. Compared to existing user modeling techniques based on collaborative filtering and deep learning, this method has significant advantages. Traditional methods require constructing a user-item rating matrix or training a complex neural network, resulting in a computational complexity of O(n log n). Or even higher, while methods based on statistical mechanics only need to calculate the probability distribution and energy function, reducing the complexity to Existing technologies suffer from significant computational overhead when handling sparse data and cold-start problems, requiring the filling of missing values ​​or extensive iterative training. A single model training session typically takes hours or even days. In contrast, statistical mechanics methods naturally address data sparsity through probability distributions, eliminating the need for additional padding or preprocessing steps. Real-time user profile calculations can be performed in milliseconds. The Maxwell-Boltzmann distribution naturally describes the equilibrium state and fluctuations of a system, effectively corresponding to the stability and variability of user interests. When a user maintains long-term interest in a particular type of content, the system tends towards equilibrium; when user interests shift, fluctuations occur. The energy function directly quantifies the attractiveness of different content to users, avoiding complex feature engineering and model parameter tuning processes. Furthermore, statistical mechanics methods offer excellent interpretability. User interest intensity directly corresponds to the energy function value, and changes in behavioral patterns correspond to state transition probabilities. This physically clear modeling approach facilitates understanding and application by business personnel, overcoming the interpretability limitations of deep learning black-box models and resolving the balance between computational efficiency and accuracy in user profile modeling.

[0042] The specific implementation methods of the above steps are described in detail below.

[0043] The specific implementation of step S01 involves first establishing a real-time data collection pipeline through a user behavior collector module. This collector, based on an event-driven architecture, is capable of capturing all micro-level user interactions on the platform interface. The data collection process includes pixel-level precise positioning of click coordinates, recording the user's dwell time on each page element to the millisecond level, tracking the user's finger or mouse swipe path with a sampling frequency of 100Hz, listening to and encoding the user's text input in real time, and counting the frequency of user voice interactions, including the number of voice inputs and voice command triggers. The collected raw data requires preprocessing. First, the micro-interaction data is normalized to eliminate dimensional differences between different data dimensions. A minimum-maximum standardization method is used to map all values ​​to the 0-1 range. Then, the micro-interaction data is categorized and labeled according to behavior type, establishing a classification system that includes behavior type identifiers such as click, browsing, input, and voice. Setting the time window is crucial for constructing state vectors. This implementation uses a sliding time window mechanism with a window length of 5 minutes and a sliding step of 1 minute. Within each time window, multiple behavior type identifiers are combined into a multi-dimensional state vector. The dimension of the state vector is determined based on the number of behavior types, typically ranging from 8 to 16 dimensions. Finally, the state vectors are arranged in chronological order to form a set of micro-state vectors representing user behavior. This set provides the foundational data for subsequent modeling and analysis.

[0044] The specific implementation of step S02 is based on the Maxwell-Boltzmann distribution principle in statistical mechanics to establish a probabilistic model of user behavior states. This principle can effectively describe the statistical laws of particle energy distribution in the system. First, the frequency of occurrence of each behavior dimension in the set of micro-state vectors of user behavior is statistically analyzed, and the probability distribution of each behavior state is calculated. A probability matrix of behavior states is established using frequency statistics. Based on the mathematical form of the Maxwell-Boltzmann distribution, user behavior states are analogized to the energy states of gas molecules. Each behavior state corresponds to an energy value; high-frequency behavior states are assigned lower energy values, and low-frequency behavior states are assigned higher energy values. This mapping relationship establishes a user behavior energy spectrum. The construction process of the energy function includes weighted fusion of behavior state energy values ​​and content semantic features. The weight coefficient is determined according to the importance of the content type, with a value range of 0.1 to 0.9. To reflect the timeliness of user interests, an exponential decay function is used to adjust the time weight of historical behavior data. The decay coefficient is set to 0.95, with the time unit being days, so that the weight of behavior data further away from the current time is smaller. The above steps yield the final user interest energy function, which quantifies the attractiveness of different content types to users, providing a scientific basis for subsequent personalized recommendations and content matching.

[0045] The specific implementation of step S03 involves using a sparse matrix data structure to represent the relationships between users and content entities. This matrix is ​​designed according to the efficiency principles of large-scale data processing. The row dimension of the matrix represents macro-level user group classification, dividing users into different groups based on their basic attributes and behavioral characteristics. The number of groups is typically set to 50 to 200. The column dimension of the matrix represents micro-level content entities, including content units in various media formats such as articles, videos, audio, and images. The number of content entities can reach millions. The matrix element values ​​represent the association strength weight between user groups and content entities. The weight values ​​are calculated based on the user's historical interaction behavior, including a comprehensive score of indicators such as click-through rate, dwell time, and number of shares. The weight values ​​range from 0 to 1. Due to the obvious sparsity of the relationships between users and content entities, most users only interact with a small amount of content; therefore, the proportion of non-zero elements in the matrix is ​​usually less than 5%. To address the computational complexity issue caused by high-dimensional sparse data, a singular value decomposition algorithm is used to reduce the dimensionality of the user-content association matrix, projecting the original high-dimensional feature space onto a low-dimensional semantic feature space. The dimensionality reduction process employs alternating least squares iterative optimization, a method that effectively addresses objective function optimization problems in matrix factorization. The number of iterations is set to 100 to 500, and the convergence threshold is set to... To prevent overfitting and preserve the sparsity of the data, a regularization term is introduced into the loss function, with the regularization parameter set between 0.01 and 0.1. The user feature matrix and content feature matrix obtained from the decomposition form the basic representation of the low-dimensional semantic feature space, which typically has a dimension one-tenth that of the original space, effectively reducing the complexity of subsequent calculations.

[0046] The specific implementation of step S04 involves using a user profile semantic enhancement model to perform deep clustering analysis on the sparse matrix. This model, based on deep learning technology, is designed to identify complex user behavior patterns. First, the similarity distance between users is calculated based on the user feature vectors obtained in step S03, using cosine similarity as the distance metric. This metric effectively captures the angular similarity between high-dimensional vectors, with a computational complexity of linearity. The construction of the similarity matrix provides the foundational data for subsequent clustering analysis. The element values ​​in the matrix represent the degree of behavioral similarity between user pairs, ranging from -1 to +1. The clustering algorithm employs density peak clustering, which, compared to traditional K-means clustering, offers better noise robustness and cluster shape adaptability, automatically determining the number of cluster centers without pre-setting. The core idea of ​​density peak clustering is to find data points with high local density that are far from other high-density points as cluster centers. The density threshold is set to 2% to 5% of the total data volume, and the distance threshold is set to 1.5 times the average distance. During the clustering process, a center vector and variance parameter are calculated for each user micro-cluster unit. The center vector represents the typical behavioral characteristics of the cluster, and the variance parameter reflects the degree of dispersion within the cluster. A variance threshold of 0.3 is set as an evaluation criterion for clustering quality. The attention weight parameters of the user profile semantic enhancement model are dynamically adjusted based on three key parameters: user historical interaction frequency, content semantic similarity, and time decay factor. The weights for historical interaction frequency, content semantic similarity, and time decay factor are set to 0.4 and 0.2 respectively. Finally, a mapping relationship between cluster units and interest tags is established, forming a micro-user profile template that includes interest weight distribution and behavioral sequence features.

[0047] The specific implementation of step S05 is to construct a user-specific semantic graph structure based on knowledge graph triple extraction technology. This technology can extract structured semantic relationships from unstructured user behavior content. First, named entities and relationship pairs are extracted from user interaction content. An entity recognition method combining rules and statistics is used. The rule base includes recognition patterns for common entity types such as person names, place names, organization names, and product names. The statistical method uses a conditional random field model for sequence labeling, with an accuracy threshold set above 85%. Entity relationship extraction employs a remote supervised learning method, utilizing an existing knowledge base as an automatic annotation source for training data, reducing the cost and workload of manual annotation. Semantic similarity between entities is calculated based on a pre-trained language model. A BERT model based on the Transformer architecture is used as the semantic encoder. This model can capture context-related semantic representations. Similarity calculation uses a cosine distance metric, with a threshold set to 0.6 as the criterion for semantic association. The construction of the user-specific semantic graph structure uses the user as the central node, establishing a multi-hop semantic relationship graph. The graph depth is set to 3 hops to balance computational complexity and semantic coverage. The graph structure includes four types of nodes: user nodes, content nodes, topic nodes, and tag nodes. Each node type has different attribute characteristics and connection patterns. User nodes store basic user attributes and behavioral preference information, content nodes store semantic features and statistical attributes of content, topic nodes represent abstract domain of interest concepts, and tag nodes represent specific attribute labels. Semantic type labels and weight attributes are assigned to each node in the graph. The weight values ​​are calculated based on the node's importance and centrality, using the PageRank algorithm for weight allocation. The number of iterations is set to 20, and the convergence threshold is set to... Node representation learning is performed using a graph neural network, employing a graph convolutional network architecture. This network can aggregate information from neighboring nodes and update the node's representation vector. The network has 3 layers, 128 hidden layer dimensions, a learning rate of 0.001, and 100 training epochs.

[0048] The specific implementation of step S06 involves encapsulating the user profile modeling service and knowledge graph construction service using the MCP protocol to achieve standardized communication and data exchange between systems. First, a standardized user profile data interface specification is defined, adhering to RESTful API design principles and including detailed descriptions of data format, transmission protocol, and error handling. The data format uses a JSON structure, supports nested objects and array types, and field naming follows camelCase. Data compression uses the gzip algorithm to reduce network transmission overhead. The user behavior micro-state vector set and macro-statistical features are encapsulated into a unified message format. The message header includes meta-information such as version number, timestamp, and data type, while the message body contains the specific user profile data content. An asynchronous communication mechanism is established between services to support real-time data exchange. Message queue technology is used to decouple services, with a message queue capacity of 10,000 messages and a message lifetime of 24 hours. This asynchronous processing mechanism effectively addresses latency differences and load fluctuations between services, improving the overall system response performance. System stability is ensured by allocating computing resources through a load balancing algorithm. A combination of round-robin and weighted round-robin scheduling is employed, dynamically adjusting weight allocation based on server CPU utilization, memory usage, network bandwidth, and other metrics. The load balancer features a health check mechanism with a 30-second check interval and a 5-second timeout. Three consecutive failures mark the server as unavailable. The system monitoring module tracks the operational status and performance metrics of each service in real time. When CPU utilization exceeds 80% or memory utilization exceeds 85%, an automatic scaling mechanism is triggered, with the scaling ratio set at 50% of the current resources.

[0049] The specific implementation of step S07 involves establishing a dynamic user profile update mechanism. This mechanism can adjust the user interest energy function parameters and individual user semantic graph structure weights in real time based on newly added behavioral data, maintaining the timeliness and accuracy of the profile model through a time window sliding strategy. First, an incremental update trigger with a fixed time interval is set. The trigger interval is determined based on the platform's data update frequency: 15 minutes for high-frequency interaction platforms and 1 hour for low-frequency interaction platforms. The trigger is implemented using a scheduled task framework, supporting distributed deployment and failover. The task execution timeout is set to half the trigger interval to ensure timely task completion. The impact weight of newly added behavioral data on the existing profile model is calculated. Information gain theory is used to quantify the value contribution of new data. The calculation of the impact weight considers three dimensions: novelty, importance, and relevance of the data. The weight threshold is set to 0.05; data below this threshold is considered noise and filtered out. An online learning algorithm is used to update the user interest energy function parameters. This algorithm has better real-time performance and memory efficiency compared to batch learning. An adaptive adjustment strategy is used for the learning rate, with an initial learning rate of 0.01, decreasing by 0.1% after each update. The system updates the node and edge weights of the user's individual semantic graph structure through graph structure modification operations. The graph update algorithm employs an incremental approach, recalculating only the changed local subgraphs to avoid the computational overhead of full graph reconstruction. Node weight updates use a moving average method: the new weight equals the historical weight multiplied by a decay coefficient, plus the newly calculated weight multiplied by an update coefficient. The decay coefficient is set to 0.9, and the update coefficient to 0.1. Edge weight updates consider changes in the weights of connected nodes and changes in interaction frequency, using a weighted fusion method to calculate the new edge weight values. The system records the history of profile changes to support model rollback and version management. Historical records use a snapshot storage mechanism, generating a complete snapshot daily for 30 days. Snapshot data uses differential compression technology to reduce storage space usage. The version management system supports model version comparison, rollback operations, and A / B testing, providing technical support for model optimization and performance evaluation.

[0050] The key technical ideas of this invention include three core innovations: user behavior modeling based on statistical mechanics, construction of a multi-level semantic graph structure, and a dynamic adaptive update mechanism. The user behavior modeling approach based on statistical mechanics introduces the Maxwell-Boltzmann distribution principle from physics into the field of user profiling. By analogy of user micro-behavior to the random motion of gas molecules, a mathematical mapping relationship is established from individual behavior to group statistical characteristics. Compared to existing collaborative filtering and deep learning methods, the statistical mechanics method has advantages such as solid theoretical foundation, low computational complexity, and strong interpretability. It can naturally handle data sparsity without complex feature engineering and model parameter tuning processes. Furthermore, it directly quantifies the attractiveness of content to users through energy functions, avoiding the black-box modeling problem of traditional methods. The multi-level semantic graph structure construction approach transforms user behavior from discrete interaction records into a structured semantic relationship network using knowledge graph technology. Compared to traditional vectorized representation methods, graph structures can better maintain the integrity and hierarchy of semantic relationships, support multi-hop reasoning and association discovery, and provide richer information dimensions for complex user interest analysis. The dynamic adaptive update mechanism establishes a real-time model parameter adjustment and graph structure update strategy. Compared to static batch update methods, the dynamic mechanism can promptly capture changing trends in user interests and the impact of sudden events, maintaining the timeliness and accuracy of the user profile model. It exhibits a significant adaptive advantage, particularly in rapidly changing network environments. The synergistic effect of these three technical approaches forms a complete user profile construction system. Statistical mechanics modeling provides a scientific theoretical foundation and an efficient computational framework; semantic graph structure offers rich feature representation and reasoning capabilities; and the dynamic update mechanism ensures the model's real-time performance and accuracy. These three elements complement and promote each other, jointly addressing the shortcomings of existing technologies in terms of computational efficiency, model accuracy, and system stability, providing a complete technical solution for large-scale user profile applications.

[0051] The detailed structure of the user profile semantic enhancement model is based on the Transformer architecture, comprising four main components: a user behavior encoding layer, a semantic feature extraction layer, an attention fusion layer, and a profile generation layer. The user behavior encoding layer transforms the original set of user behavior micro-state vectors into a high-dimensional semantic embedding representation. This layer employs a multi-head self-attention mechanism to capture long-distance dependencies in the behavior sequence. The encoding layer contains 512 hidden units, with 8 attention heads, and positional encoding is generated using sine and cosine functions. The semantic feature extraction layer performs deep abstraction of the encoded behavioral features through a multilayer perceptron network. This layer contains three fully connected sublayers with hidden layer dimensions of 512, 256, and 128, respectively. The ReLU activation function is used, and the dropout ratio is set to 0.1 to prevent overfitting. The attention fusion layer is the core innovative component of the model. It employs an adaptive weighting mechanism to dynamically adjust the importance of each feature dimension based on user interaction patterns. This layer calculates a weighted fusion of user historical interaction frequency, content semantic similarity, and time decay factor. The weight parameters are dynamically learned through a gating mechanism, and the sigmoid activation function ensures that the weight values ​​are between 0 and 1. The user profile generation layer maps the fused features to the final user profile feature representation vector. This layer employs linear transformation and layer normalization operations, with an output dimension of 256, covering multiple dimensions such as user interests, behavioral patterns, and social attributes. The model's training dataset construction process includes large-scale collection of user behavior log data, sourced from platforms across various fields such as e-commerce, social media, entertainment, and education, spanning 12 months to ensure temporal diversity. The data cleaning and standardization stages remove outliers and noisy data. Outlier detection uses a statistical distribution-based method, setting a threshold of 3 times the standard deviation. Noise filtering uses a low-pass filter to remove high-frequency disturbances. User behavior sequence samples are constructed using a sliding window method, with a window length of 30 days and a sliding step of 7 days. Each sample contains the user's complete behavior sequence within the window period and the corresponding profile label. Data augmentation techniques, such as sequence perturbation, time shifting, and feature fusion, expand the diversity of training samples, with an augmentation ratio set to twice the original data. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio to ensure uniform distribution of different user groups across the datasets. The model training employs the Adam optimization algorithm for gradient descent updates, with a learning rate of 0.0001, a batch size of 64, and 200 training epochs. The learning rate decay strategy uses cosine annealing, dynamically adjusting the learning rate during training to prevent oscillations. The loss function is a weighted combination of cross-entropy loss and mean squared error loss, with a weight of 0.6 for classification and 0.4 for regression. An early stopping mechanism monitors the loss function value on the validation set; training stops if there is no improvement after 10 consecutive epochs. Model checkpoints are saved every 5 epochs to support model rollback.The fundamental reason why the user profile semantic enhancement model is suitable for solving the technical problem of this invention is that it can simultaneously process sequential behavioral data and structured semantic information, achieving effective fusion of different information sources through an attention mechanism. Compared with traditional matrix factorization methods such as singular value factorization and nonnegative matrix factorization, the semantic enhancement model can capture the temporal dependencies and semantic hierarchical structure of user behavior, avoiding the problems of information loss and semantic omission in traditional methods. Compared with existing deep learning methods such as recurrent neural networks and convolutional neural networks, the parallel computing characteristics and global attention mechanism of the Transformer architecture can better handle long sequence data and long-range dependencies, especially showing significant modeling advantages in scenarios with long user behavior sequences. The model's adaptive weight mechanism can dynamically adjust the importance of features according to the behavioral characteristics of different users, which is more flexible and accurate than static feature weight settings, effectively improving the effect and generalization ability of personalized modeling.

[0052] It should be noted that the present invention also solves the following technical problems.

[0053] The challenge of identifying users' true interests: In existing technologies, users' explicit behavioral data often fails to fully reflect their true interests and preferences. Users may exhibit behaviors inconsistent with their true interests due to curiosity, accidental actions, or social needs. Traditional methods treat all behavioral data equally, making it difficult to distinguish which behaviors truly represent a user's core interests. This results in profiles that may be biased towards superficial or temporary interest characteristics. This invention utilizes the energy minimization principle from statistical mechanics. True user interests correspond to low-energy stable states of a system. These behaviors exhibit strong clustering and persistence in terms of time and content. By calculating the energy distribution and state transition probabilities of behavioral sequences, stable behavioral patterns representing users' core interests can be identified. Combined with semantic association analysis in knowledge graphs, the internal logical consistency of the behaviors is further verified, thereby accurately extracting users' true interest characteristics and avoiding being misled by superficial behaviors.

[0054] The issue of accuracy in multi-dimensional behavioral feature fusion: User interactions on AI platforms encompass information across multiple dimensions, including click location, dwell time, swipe speed, and input content. Features in different dimensions have varying data types, value ranges, and noise levels. Traditional feature fusion methods typically employ simple linear weighting or feature concatenation strategies, which struggle to accurately model the complex nonlinear relationships between features across dimensions, easily leading to the dilution of important features or the amplification of noisy features. This invention maps multi-dimensional behavioral features uniformly onto the state space of statistical mechanics, with each dimension corresponding to a component of the state vector. The joint probability distribution of each dimension's features is described using the multivariate form of the Maxwell-Boltzmann distribution. The invention automatically learns the weight coefficients and interaction relationships of each dimension's features, achieving feature fusion based on physical principles rather than empirical weight allocation. Furthermore, by leveraging the global optimization properties of the energy function, the optimal feature combination pattern can be found in the multi-dimensional space, significantly improving the accuracy and robustness of feature fusion.

[0055] The problem of insufficient generalization ability of user behavior patterns: Existing user modeling methods often rely excessively on behavioral data from specific platforms or scenarios, resulting in user profiles that lack cross-platform and cross-domain generalization capabilities. When users exhibit behavior in different environments, the original profile model may become ineffective or its accuracy may drop significantly, limiting the practicality of user profiles in multi-scenario applications. This invention is based on the universality principles of statistical mechanics. The Maxwell-Boltzmann distribution, as a fundamental law describing thermal equilibrium systems, maintains a consistent mathematical form across different physical systems. This universality enables the user profile model constructed in this invention to have good generalization ability. By abstracting user behavior into statistical mechanical states, the specific influence of specific platforms and scenarios is eliminated, and the essential features of user interests are extracted. Combined with the semantic representation capabilities of knowledge graphs, user interest patterns can be mapped to a general conceptual space, enabling cross-platform user profile migration and application, and providing a unified technical foundation for personalized services in multiple scenarios.

[0056] Specifically, the principle of this invention is as follows: The core of this invention's ability to solve the problems of noise interference in user behavior data and the accuracy of user profiles lies in its ingenious application of the probability distribution theory in statistical mechanics, which possesses the natural ability to identify outlier data points. In statistical mechanics, when particles in a system reach thermal equilibrium, their energy distribution follows the Maxwell-Boltzmann distribution law. The energy of the vast majority of particles is concentrated in the main part of the distribution, with only a very small number of particles located in the tail region. This statistical law provides a theoretical basis for identifying abnormal states in a system. This invention models user behavior as a similar physical system. Each user interaction corresponds to an energy state. Normal interest-driven behavior, due to its inherent regularity and consistency, statistically manifests as the main part conforming to the Maxwell-Boltzmann distribution, while noisy behavior, due to its randomness and chance, appears as an anomaly in the energy distribution, deviating from the main distribution. In its specific implementation, this invention calculates the probability density and energy value of each behavioral state to establish a user interest energy function to describe the distribution characteristics of the behavioral data. Genuine user interest behaviors typically exhibit high repetition and continuity, corresponding to low energy states and high probability densities. Noise behaviors, such as accidental clicks and bot actions, are usually isolated and unrelated, corresponding to high energy states and extremely low probability densities. Using distribution theory from statistical mechanics, the system can automatically calculate the degree to which each data point deviates from the normal distribution, set appropriate confidence intervals and energy thresholds, and identify data points with excessively low probability densities or excessively high energy as noise for filtering. Furthermore, the sparse matrix representation and knowledge graph technology employed in this invention further enhance the accuracy of noise identification. By analyzing the semantic consistency and correlation rationality of user behavior data, the authenticity of the behavior can be verified at the content level. For example, a user's actions on multiple completely unrelated content categories within a short period may indicate abnormal behavior, while continuous interaction between related content is more likely to reflect genuine interests. The dynamic update mechanism continuously monitors changes in data quality and adjusts noise identification parameters and filtering strategies in a timely manner, ensuring that user profiles are always built upon high-quality behavioral data, thereby guaranteeing the accuracy and reliability of the profile results.

[0057] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0058] The specific implementation of step S01 involves acquiring micro-interaction data through a user behavior collector and preprocessing it. The micro-interaction data is then normalized using a minimum-maximum standardization method, as shown below:

[0059] ;

[0060] In the formula, For the first Normalized values ​​of each data point; For the first The original values ​​of each data point; The minimum value in the dataset; The maximum value in the dataset; This is the index for the data points, with values ​​ranging from 1 to the total number of data points. The method for obtaining these parameters is as follows: Data is acquired in real time through a user behavior collector, including click coordinates, dwell time, and swipe trajectory data. and The user behavior micro-state vector is calculated by traversing all data points within the current time window. The representation of the user behavior micro-state vector is as follows:

[0061] ;

[0062] In the formula, For time The state vector; For time No. The identifier value of the behavior type, ; This represents the total number of behavior types, typically ranging from 8 to 16. This is a time index, in minutes.

[0063] The specific implementation of step S02 is based on constructing a user interest energy function according to the Maxwell-Boltzmann distribution principle. The probability distribution of user behavior states is represented as follows:

[0064] ;

[0065] In the formula, Behavioral state The probability is dimensionless and ranges from 0 to 1; Behavioral state The corresponding energy value is expressed in joules. It is a normalization constant, dimensionless; is the Boltzmann constant, which takes the value of 0.1 in this system, and its unit is joules per kelvin; This is the system temperature parameter, with a value of 1.0, and the unit is Kelvin; This is the behavior state index, with a value ranging from 1 to the total number of behavior states. The parameter is obtained as follows: Based on behavior frequency Calculations show that ,in For the first The frequency of each behavioral state The total frequency of all behaviors; The result is obtained by summing and normalizing the probabilities of all states. ,in This represents the total number of behavioral states. The user interest energy function is constructed as follows:

[0066] ;

[0067] In the formula, For content The user interest energy function value, in joules; For indexing or identifying content objects; For the first The weight coefficient of each behavioral state is dimensionless and ranges from 0 to 1. For the first The energy value of a behavioral state, in joules; For content The semantic feature value is dimensionless and ranges from 0 to 1; This is the time decay factor, with a value of 0.05, in days. ; The time difference from the current time, in days; This represents the total number of behavioral states, typically ranging from 8 to 16.

[0068] The specific implementation of step S03 involves using a sparse matrix to represent the relationship between users and content and performing matrix decomposition. The user-content relationship matrix is ​​represented as follows:

[0069] ;

[0070] In the formula, User content association matrix; For the first The user group and the first The association strength weight of each content entity; For the number of users; For the number of content entities; Indexed for user groups, with values ​​ranging from 1 to... ; For content entity indexes, values ​​range from 1 to... The singular value decomposition is represented as follows:

[0071] ;

[0072] In the formula, for The left singular matrix; for a diagonal matrix; for The right singular matrix; For matrix The transpose of . The loss function of alternating least squares is expressed as follows:

[0073] ;

[0074] In the formula, The loss function; For user groups eigenvectors; For content eigenvectors; This is the regularization parameter, with a value ranging from 0.01 to 0.1; and These are the squares of the 2-norm of the eigenvectors.

[0075] The specific implementation of step S04 involves using a user profile semantic enhancement model for cluster analysis. The cosine similarity calculation between users is represented as follows:

[0076] ;

[0077] In the formula, For users and users The cosine similarity is dimensionless and ranges from -1 to 1. It is the dot product of two user feature vectors; For users eigenvectors; For users eigenvectors; and These are the L2 norms of the user feature vectors; Indexed for users. The local density calculation in density peak clustering is represented as follows:

[0078] ;

[0079] In the formula, For data points The local density is dimensionless; For data points and The distance between them; To cut off the distance; For indicator functions, when hour ,otherwise ; Index the current data point; Indexing neighboring data points. The dynamic adjustment of attention weights is represented as follows:

[0080] ;

[0081] In the formula, For dynamic attention weights; It is the sigmoid activation function. ; This is the weight parameter matrix; Historical interaction frequency characteristics; For content semantic similarity features; It is a time decay characteristic; This is a bias term.

[0082] The specific implementation of step S05 is based on constructing a semantic graph structure for individual users using knowledge graph technology. The semantic similarity calculation between entities is represented as follows:

[0083] ;

[0084] In the formula, For entities and The semantic similarity is dimensionless and ranges from -1 to 1. and Entities and The semantic embedding vectors typically have dimensions ranging from 128 to 512. The PageRank weighting algorithm is represented as follows:

[0085] ;

[0086] In the formula, For nodes PageRank value, dimensionless; The damping coefficient is 0.85 and is dimensionless. This represents the total number of nodes in the graph; Pointing to a node The set of nodes; For nodes The number of out-degrees.

[0087] The specific implementation method of step S06 is the same as described above, and will not be repeated in detail here.

[0088] The specific implementation of step S07 is to establish a dynamic update mechanism, and the impact weight calculation of the newly added data is expressed as follows:

[0089] ;

[0090] In the formula, The weights for the impact of newly added data are dimensionless. Information gain for newly added data; Assign a relevance score to the newly added data, with a value ranging from 0 to 1; The timeliness score for newly added data ranges from 0 to 1. These are weighting coefficients, with values ​​of 0.4, 0.4, and 0.2 respectively, and are dimensionless. This is a newly added dataset. The information gain is calculated as follows:

[0091] ;

[0092] In the formula, The entropy value of the original dataset; This is the original dataset; Add new data attributes; For attributes All possible values; The attribute value is A subset of data; This is the size of the original dataset; Size of the data subset; For attributes The specific value to be taken; For data subset The entropy value. The parameter updates for online learning are represented as follows:

[0093] ;

[0094] In the formula, These are the updated model parameters; These are the current model parameters; The learning rate; This represents the gradient of the loss function with respect to the parameters. Input data; For label data; The time step index is used. The graph structure weight update is represented using the moving average method as follows:

[0095] ;

[0096] In the formula, The updated weights; Historical weighting; For the newly calculated weights; The attenuation coefficient is 0.9.

[0097] The principles and effects of each formula and mathematical model are explained below. Normalization formula The principle is based on linear transformation to map data of different dimensions to a unified interval, eliminating the impact of data scale differences on subsequent calculations. Compared with the traditional Z-score standardization method, it has the advantage of maintaining the original data distribution shape, avoiding the excessive influence of outliers on the standardization results, and improving the stability and reliability of data preprocessing.

[0098] Maxwell-Boltzmann distribution formula The principle originates from the energy distribution laws of particles in statistical physics, where the exponential term... The normalization factor represents the negative exponential relationship between energy and probability. By ensuring the normalization of probability distribution, user behavior states are analogized to molecular energy states, and the negative correlation between state probability and energy is described by an exponential function. Compared with traditional frequency statistics methods, this approach can better describe the randomness and regularity of user behavior, establishing a theoretical bridge from micro-level individual behavior to macro-level group characteristics, and providing a solid mathematical foundation for user profile modeling.

[0099] User Interest Energy Function The study comprehensively considered three key factors: behavioral state energy, semantic features, and time decay, among which the weighted summation term... This reflects the difference in importance between different behavioral states, semantic feature terms It introduces deep semantic information of the content and an exponential decay term. It simulates the natural decay process of user interest over time. Compared with static interest modeling methods, it can dynamically capture the time-varying characteristics of user interests, significantly improving the accuracy and timeliness of personalized recommendations.

[0100] Singular Value Decomposition Formula The principle is based on the linear algebraic decomposition theory of matrices, which decomposes a high-dimensional sparse matrix into the product of three low-dimensional matrices. By retaining the main singular values, it achieves data dimensionality reduction and noise removal. Compared with the traditional principal component analysis method, it has better numerical stability and computational efficiency when processing sparse data, and effectively solves the storage and computation problems of large-scale user content matrices.

[0101] Alternating Least Squares Loss Function The principle combines minimizing reconstruction error with regularization constraints, with the squared error term in the first half... To ensure the decomposition results can accurately reconstruct the original matrix, the regularization term in the latter half... To prevent parameter overfitting and maintain the sparsity of the solution, it has better generalization ability and numerical stability than unconstrained matrix factorization methods, and can still obtain reliable decomposition results in sparse data scenarios.

[0102] Cosine similarity formula The principle is based on the geometric concept of the angle between vectors. It measures the directional similarity of two vectors by calculating the standardized inner product of the vectors. It is not affected by the length of the vectors and only focuses on the directional features. Compared with other similarity measurement methods such as Euclidean distance, it has better distinguishing ability and computational stability in high-dimensional sparse data, and provides a reliable similarity judgment standard for users' clustering analysis.

[0103] Local density formula in density peak clustering Indicator functions were used The statistical counting principle, when the distance Less than the cutoff distance It contributes 1 count when the count is active and 0 count otherwise. The density of the neighbors around each data point is obtained by accumulating the statistics. Compared with the density estimation method based on kernel function, it has better computational efficiency and parameter robustness.

[0104] Dynamic attention weight formula The principle employs the nonlinear mapping mechanism of neural networks, where linear combination terms... Integrating multidimensional feature information, the sigmoid activation function By ensuring that the weight values ​​are within the range of 0 to 1, the importance of each feature dimension is adaptively adjusted by learning linear combinations and nonlinear transformations of different features. Compared with the fixed-weight feature fusion method, it can dynamically adjust the importance of features according to the behavior patterns of different users, which significantly improves the personalized modeling ability and adaptability of the model.

[0105] PageRank formula The principle is based on random walks and Markov chain theory, where random jump terms... The weight propagation term represents the probability of randomly jumping from any node to the current node. It represents the importance weight propagated from neighboring nodes. By iteratively calculating the importance propagation process of nodes, the damping coefficient balances the influence of local connectivity and global uniform distribution. Compared with simple degree centrality and other graph analysis methods, it can more accurately identify key nodes in the graph, providing a scientific theoretical basis for evaluating the importance of entities in user semantic graphs.

[0106] Information gain formula The principle is based on the concept of entropy in information theory, where the initial entropy... The conditional entropy term represents the uncertainty of the dataset. This represents the weighted average entropy after segmentation by attribute. The difference between the two values ​​quantifies the contribution of the new attribute to the data classification. Compared with simple data volume statistics methods, it can more accurately assess the value of newly added data and provides a scientific data screening standard for dynamic update mechanisms.

[0107] Online learning to update formulas The principle is based on gradient descent optimization theory, which calculates the gradient direction of the loss function with respect to the parameters. Gradually adjust model parameters, learning rate By controlling the step size of parameter updates, it can respond to changes in new data in real time compared to batch learning methods. While maintaining model performance, it significantly reduces the consumption of computing resources and storage requirements, providing an efficient update mechanism for large-scale online user profiling systems.

[0108] Moving average weight update formula The principle is based on the exponentially weighted moving average theory, achieving smooth updates through a linear combination of historical weights and newly calculated weights, with a decay coefficient. By controlling the relative importance of historical and new information, it has better numerical stability and noise resistance than the direct replacement update method, avoids the excessive impact of single abnormal data on model performance, and ensures the continuity and consistency of user profiles.

[0109] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team is responsible for building an accurate user profile system for an intelligent content recommendation platform. This platform has 1.5 million daily active users, covering multiple content areas such as news, video, music, and shopping. Traditional collaborative filtering and simple tagging methods can no longer meet the rapidly changing personalized needs of users. The technical team decided to adopt a user profile construction method based on statistical mechanics principles to solve this problem.

[0110] The technical team first deployed a user behavior collector to gather real-time micro-interaction data from users on the platform. The collected data included click coordinates (accurate to the pixel level), page dwell time (in milliseconds), swipe trajectories (recording swipe speed and direction), input content (including search keywords and comment text), and voice interaction frequency. Within 24 hours, the system collected 280 million user behavior data records. The technical team normalized this micro-interaction data, eliminating dimensional differences between different data dimensions. Click coordinates were normalized to between 0 and 1, dwell time was logarithmically transformed, and swipe trajectories were converted into standardized vector representations. The data was then categorized and labeled according to behavior type, forming six basic behavior type identifiers: click, browse, search, share, favorite, and comment. The technical team set a 5-minute time window, combining multiple behavior type identifiers within the same window into state vectors. Each state vector contained 128 dimensions of features. These state vectors were arranged chronologically to form a set of user behavior micro-state vectors, generating a total of 4.2 × A state vector.

[0111] In the phase of building the probabilistic model of user behavior states, the technical team established a mapping relationship from micro-behaviors to macro-statistical characteristics based on the Maxwell-Boltzmann distribution principle. First, they statistically analyzed the frequency of occurrence of each behavioral dimension and calculated the probability distribution function, discovering that user click behavior follows a typical long-tail distribution, with 20% of popular content receiving 78% of the clicks. The technical team then calculated the energy value of each behavioral state using the Maxwell-Boltzmann distribution formula E = -kT×ln(P), where k is the Boltzmann constant (set to 1.38×). T represents the system temperature parameter (dynamically adjusted between 0.1 and 2.0), and P represents the state probability. High-frequency behaviors correspond to lower energy values; for example, news content frequently viewed by users has an energy value of 0.15, while occasionally viewed niche music content has an energy value of 2.34. The technical team weighted and fused the behavioral state energy values ​​with content semantic features, using a pre-trained BERT model to extract content semantic embedding vectors with a dimension of 768. This was achieved through an exponential decay function. Historical behaviors are weighted over time, with λ set to 0.05 and t representing the time interval in days, ultimately establishing a user interest energy function.

[0112] The technical team used a sparse matrix to represent the relationships between users and content entities. The matrix size was 1.5 million × 500,000, where rows represented users and columns represented content entities, achieving a sparsity of 99.7%. The matrix element values ​​represented the weight of the user's association with the content, calculated using the following formula: ,in Let i be the frequency of user i's interaction with content j. and Let U represent the total number of interactions between user i and content j, respectively. The technical team used singular value decomposition (SVD) to reduce the dimensionality of the sparse matrix, projecting the original 1.5 million-dimensional user behavior space onto a 200-dimensional low-dimensional semantic feature space. The loss function of matrix decomposition was iteratively optimized using alternating least squares, with a regularization parameter of 0.01 set to prevent overfitting and maintain sparsity. After 32 iterations, the reconstruction error converged to 0.0156, yielding the user feature matrix U and content feature matrix V as the basic representation of the low-dimensional semantic feature space.

[0113] In building the user profile semantic enhancement model, the technical team designed a multi-head attention network based on the Transformer architecture. The model comprises four main components: a user behavior encoding layer, a semantic feature extraction layer, an attention fusion layer, and a profile generation layer. The user behavior encoding layer uses a 6-layer Transformer encoder, with 8 attention heads per layer and a hidden layer dimension of 512. The semantic feature extraction layer integrates three sub-modules: named entity recognition, relation extraction, and semantic embedding, enabling the identification of entity types and semantic relationships from user interaction content. The attention fusion layer employs an adaptive weighting mechanism, dynamically adjusting the importance of each feature dimension based on three parameters: user historical interaction frequency, content semantic similarity, and time decay factor. The technical team trained the model on a dataset containing 2.8 million user behavior samples, using the Adam optimization algorithm with a learning rate of 0.001 and a batch size of 256. After 45 training epochs, the model achieved an accuracy of 87.3% on the validation set, and the loss function converged to 0.234.

[0114] The technical team utilized a trained user profile semantic enhancement model to perform cluster analysis on the sparse matrix, identifying micro-clusters of users with similar behavioral patterns. Based on the 200-dimensional user feature vectors, a cosine similarity distance matrix was calculated, and density peak clustering was used to divide 1.5 million users into 2847 clusters. The average number of users in each cluster was 527, with a standard deviation of 89. The team calculated the center vector and variance parameters for each cluster, establishing a mapping relationship between clusters and interest tags. Table 1 shows the distribution of interest preferences corresponding to different user clusters.

[0115] Table 1. Distribution of User Clustering Unit Interest Preferences

[0116]

[0117] During the knowledge graph construction phase, the technical team extracted entity relationship pairs from user behavior content using triplet extraction technology. The system automatically identified 1.78 million entity nodes, including four types: user nodes, content nodes, topic nodes, and tag nodes. User nodes accounted for 84.2% of the total number of nodes, content nodes for 12.7%, topic nodes for 2.8%, and tag nodes for 0.3%. The team constructed a user-individual semantic graph structure containing 3.45 million edges, with an average of 1.94 edges connected to each user. A pre-trained RoBERTa model was used to calculate the semantic similarity between entities, constructing a user-centric multi-hop semantic relationship graph. Semantic type labels and weight attributes were assigned to each node in the graph, and node representation learning was performed using a graph neural network, ultimately obtaining a 256-dimensional node embedding vector.

[0118] The technical team encapsulated user profile modeling and knowledge graph construction services using the MCP protocol, enabling the fusion of micro-level state vector sets and macro-level profile features. A standardized user profile data interface specification was defined, supporting both JSON and Protocol Buffers data formats. The user behavior micro-level state vector sets and macro-level statistical features were encapsulated into a unified message format, with the message header containing metadata such as user ID, timestamp, and data version. An asynchronous communication mechanism based on Apache Kafka was established, supporting the processing of 15,000 user profile update requests per second. Load balancing was achieved through a consistent hashing algorithm, distributing computational tasks across 32 server nodes to ensure system stability and high availability.

[0119] During the implementation of the dynamic user profile update mechanism, the technical team set up an incremental updater that triggers every 5 minutes. The system monitors new behavioral data in real time and calculates its impact weight on the existing profile model. When the weight of new user behavioral data exceeds the threshold of 0.15, the profile update process is triggered. The technical team uses an online learning algorithm to update the user interest energy function parameters, with a learning rate set to 0.01 and a momentum parameter set to 0.9. The system updates the node and edge weights of the individual user's semantic graph structure through graph structure modification operations, supporting node addition, deletion, and weight adjustment operations. The system records historical versions of each profile change, supports model rollback and version management, and retains historical data for 30 days. Table 2 shows the performance statistics of the dynamic update mechanism.

[0120] Table 2 Performance Indicators of Dynamic Update Mechanism

[0121]

[0122] To verify the invention's ability to resist interference from noisy data, the technical team designed a comparative experiment. Different levels of noise were artificially injected into the original user behavior data, including accidental clicks (accounting for 12.4% of total clicks), abnormal dwell times (data exceeding three times the standard deviation of the normal range accounted for 8.7%), and invalid swipe trajectories (operations with a swipe distance of less than 5 pixels accounted for 15.2%). Table 3 shows the comparison results of the accuracy of user profiling using various methods under different noise levels.

[0123] Table 3. Comparison of User Profile Accuracy in Noisy Environments

[0124]

[0125] The technical team further analyzed the impact of noise data on user interest identification. A probabilistic model built using the Maxwell-Boltzmann distribution principle can automatically identify and filter abnormal behavior patterns, attenuating high-energy states corresponding to noise behavior. In a test environment with a 25% noise level, the system successfully identified 89.3% of the noise data, with a false positive rate of only 3.7%. Table 4 shows the statistical results of identification and processing of different types of noise data.

[0126] Table 4. Results of Noise Data Identification and Processing

[0127]

[0128] In summary, this embodiment analogizes user micro-behavior to the movement of gas molecules and effectively filters out noise interference in behavioral data by utilizing the statistical characteristics of probability distribution. This significantly improves the accuracy and stability of user profiles in noisy environments and successfully solves the technical problem in the prior art where user micro-behavioral data is easily interfered with by noise, resulting in insufficient accuracy of the user profiles.

[0129] It should be noted that the variables involved in this invention are explained in detail in Tables 5 and 6.

[0130] Table 5. Variable Explanation Table (Part 1)

[0131]

[0132] Table 6. Variable Explanation Table (Part Two)

[0133]

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing user profiles on an AI platform, characterized in that, include: The system acquires real-time micro-interaction data of users on the platform through a user behavior collector, and arranges the micro-interaction data in time series to form a set of user behavior micro-state vectors. Based on the Maxwell-Boltzmann distribution principle in statistical mechanics, a user behavior state probability model is constructed, mapping the set of user behavior micro-state vectors to a macro-statistical feature space. The user interest energy function is established by calculating the distribution entropy value of each behavior dimension. A sparse matrix is ​​used to represent the relationship between users and content entities, and a matrix factorization algorithm is used to project the high-dimensional user behavior space onto a low-dimensional semantic feature space. Using a user profile semantic enhancement model, cluster analysis is performed on sparse matrices to identify micro-clustering units of users with similar behavioral patterns; based on knowledge graph triple extraction technology, entity relationship pairs are extracted from user behavior content to construct individual user semantic graph structures; user profile modeling services and knowledge graph construction services are encapsulated through the MCP protocol to achieve the fusion processing of user behavior micro-state vector sets and macro-profile features. Establish a dynamic user profile update mechanism to adjust the parameters of the user interest energy function and the weights of the individual user semantic graph structure in real time based on newly added behavioral data, thereby completing the construction of user profiles.

2. The user profile construction method under the AI ​​platform according to claim 1, characterized in that, Micro-interaction data includes click coordinates, dwell time, swipe trajectory, input content, and voice interaction frequency. The steps to establish the user behavior micro-state vector set are as follows: normalize the micro-interaction data to eliminate dimensional differences, classify and label the micro-interaction data according to behavior type to form behavior type identifiers, combine multiple behavior type identifiers within the same time window into a state vector, and arrange the state vectors in chronological order to form the user behavior micro-state vector set.

3. The user profile construction method under the AI ​​platform according to claim 2, characterized in that, The user interest energy function is used to quantify the attractiveness of different content types to users. The specific steps for constructing the user interest energy function are as follows: calculate the probability distribution of the occurrence frequency of each behavioral dimension, calculate the energy value of each behavioral state based on the Maxwell-Boltzmann distribution formula, weight and fuse the energy value of the behavioral state with the semantic features of the content, and adjust the time weight of historical behaviors through an exponential decay function to obtain the final user interest energy function.

4. The user profile construction method under the AI ​​platform according to claim 3, characterized in that, The rows of the sparse matrix represent macro-level user group classifications, the columns represent micro-level content entities, and the matrix element values ​​represent association strength weights. The optimization steps of sparse matrix decomposition specifically involve using the singular value decomposition algorithm to reduce the dimensionality of the user-content association matrix, iteratively optimizing the loss function of matrix decomposition using alternating least squares, setting regularization parameters to prevent overfitting and maintain sparsity, and using the decomposed user feature matrix and content feature matrix as the basic representation of the low-dimensional semantic feature space.

5. The user profile construction method under the AI ​​platform according to claim 4, characterized in that, Each user micro-cluster unit corresponds to a combination of interest preferences. The specific steps for identifying user micro-cluster units are as follows: calculate the cosine similarity distance matrix based on the user feature vector, use the density peak clustering algorithm to identify user groups with similar behavioral patterns, calculate the center vector and variance parameter for each cluster unit, and establish the mapping relationship between cluster units and interest tags to form a micro-user profile template.

6. The user profile construction method under the AI ​​platform according to claim 5, characterized in that, The attention weight parameters of the user profile semantic enhancement model are dynamically adjusted based on three parameters: user historical interaction frequency, content semantic similarity, and time decay factor. The user profile semantic enhancement model is a multi-head attention network based on the Transformer architecture, which includes four main components: user behavior encoding layer, semantic feature extraction layer, attention fusion layer, and profile generation layer. The attention fusion layer adopts an adaptive weight mechanism to dynamically adjust the importance of each feature dimension according to the user interaction pattern.

7. The user profile construction method under the AI ​​platform according to claim 6, characterized in that, The user-individual semantic graph structure includes four types: user nodes, content nodes, topic nodes, and tag nodes. The specific steps for constructing the user-individual semantic graph structure are as follows: extract named entities and relation pairs from user interaction content, calculate the semantic similarity between entities based on a pre-trained language model, construct a user-centric multi-hop semantic relation graph, assign semantic type labels and weight attributes to each node in the graph, and obtain the final user-individual semantic graph structure through node representation learning using a graph neural network.

8. The user profile construction method under the AI ​​platform according to claim 7, characterized in that, The MCP protocol encapsulates and processes user behavior micro-state vector sets into macro-knowledge representations through semantic relationship links. It also jointly optimizes the individual semantic graph structure of individual users with the macro-statistical model of the global user group. Specifically, it defines standardized user profile data interface specifications, encapsulates user behavior micro-state vector sets and macro-statistical features into a unified message format, establishes an asynchronous communication mechanism between services to support real-time data exchange, and allocates computing resources through load balancing algorithms to ensure system stability.

9. The user profile construction method under the AI ​​platform according to claim 8, characterized in that, The dynamic user profile update mechanism maintains the timeliness and accuracy of the profile model through a time window sliding strategy. The specific steps of the dynamic user profile update mechanism are as follows: set incremental update triggers at fixed time intervals, calculate the impact weight of new behavioral data on the existing profile model, update the parameters of the user interest energy function using an online learning algorithm, update the node and edge weights of the individual user semantic graph structure through graph structure modification operations, and record the profile change history to support model rollback and version management.

10. The user profile construction method under the AI ​​platform according to claim 9, characterized in that, The steps for establishing the training dataset for the user profile semantic enhancement model specifically include collecting large-scale user behavior log data covering different domains and time periods, cleaning and standardizing the behavior data to remove outliers and noise, constructing user behavior sequence samples and labeling them with corresponding profile tags, expanding the diversity of training samples through data augmentation techniques, and dividing the training set, validation set, and test set to ensure the rationality of data distribution.

Images