Method and apparatus for determining user profile

Abstract

The embodiment of the present application relates to the technical field of artificial intelligence, and discloses a kind of determination method and device of user portrait, the method comprises: obtaining the behavior data of target user in current period;The behavior data of the current period is modeled and predicted to process, obtain the micro portrait of the target user in current period;Wherein, the micro portrait represents the weight distribution of the intention component of target user in current period;Based on the preset macroscopic value prediction model and the micro portrait of the current period, determine the macroscopic value index of the target user;Wherein, the macroscopic value prediction model is configured to learn the nonlinear mapping from intention component to macroscopic value index;According to the micro portrait of the current period and the macroscopic value index, obtain the user portrait of the target user.The technical scheme of the application can solve the problem of one-sidedness of user portrait in the prior art.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and apparatus for determining user profiles. Background Technology

[0002] With the widespread adoption of digital services, user profiles, as the core foundation for personalized services and value operations, directly determine service quality and long-term value creation capabilities through their modeling accuracy and functional completeness.

[0003] Currently, most mainstream user profiling methods are based on deep learning and traditional probabilistic models. They generate user profiles that represent user preferences by mining the correlation between users' historical behavior data and interest tags, and then provide related services based on these profiles. However, these methods often focus on short-term clicks and conversions as the core evaluation dimensions, only paying attention to the correlation analysis of users' explicit behaviors, while neglecting the mining and characterization of users' implicit intentions. This results in user profiles that are often too one-sided, only reflecting users' short-term preferences, and failing to deeply understand users' potential real needs and long-term value demands. Consequently, they are unable to support accurate, comprehensive, and long-term value-added service output. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide a method and apparatus for determining user profiles, which are used to solve the problem of one-sided user profiles in the prior art.

[0005] According to one aspect of the present invention, a method for determining a user profile is provided, the method comprising:

[0006] Obtain the target user's behavioral data for the current time period;

[0007] Intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period; wherein, the micro-profile represents the weight distribution of the intent components of the target user in the current time period;

[0008] Based on a preset macro-value prediction model and the micro-profile of the current time period, the macro-value indicators of the target user are determined; wherein, the macro-value prediction model is configured to learn a non-linear mapping from intent components to macro-value indicators.

[0009] Based on the micro-profile of the current time period and the macro-value indicators, the user profile of the target user is obtained.

[0010] According to another aspect of the present invention, a user profile determination apparatus is provided, comprising:

[0011] The acquisition module is used to acquire the target user's behavioral data in the current time period;

[0012] The processing module is used to perform intent modeling and prediction processing on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period; wherein, the micro-profile represents the weight distribution of the intent components of the target user in the current time period;

[0013] The determination module is used to determine the macro value index of the target user based on a preset macro value prediction model and the micro profile of the current time period; wherein the macro value prediction model is configured to learn a nonlinear mapping from intent components to macro value indexes.

[0014] The integration module is used to obtain the user profile of the target user based on the micro profile of the current time period and the macro value indicators.

[0015] According to another aspect of the present invention, an electronic device is provided, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0016] The memory is used to store at least one executable instruction that causes the processor to perform the operation of the user profile determination method described above.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing at least one executable instruction that causes an electronic device / user profile determination device to perform the following operations:

[0018] Obtain the target user's behavioral data for the current time period;

[0019] Intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period; wherein, the micro-profile represents the weight distribution of the intent components of the target user in the current time period;

[0020] Based on a preset macro-value prediction model and the micro-profile of the current time period, the macro-value indicators of the target user are determined; wherein, the macro-value prediction model is configured to learn a non-linear mapping from intent components to macro-value indicators.

[0021] Based on the micro-profile of the current time period and the macro-value indicators, the user profile of the target user is obtained.

[0022] This invention, after acquiring the target user's behavioral data for the current time period, employs quantum intent modeling combined with time-series prediction to refine and calculate the user's implicit intent, resulting in a micro-profile. Relying on a pre-trained nonlinear macro-value prediction model, the micro-intention features are converted into long-term macro-value indicators. By integrating the micro-intentions and macro-value indicators, a comprehensive user profile is ultimately formed. This approach enables precise quantitative characterization of the user's ambiguous implicit intent from the ground up, while introducing a long-term macro-value measurement dimension. It overcomes the one-sidedness of traditional user profiling, which only focuses on explicit behavior and short-term value, significantly improving the comprehensiveness and accuracy of the user profile. This allows the profile to truly reflect the user's deep needs and long-term value attributes, providing more reliable data support and decision-making basis for subsequent user-aware precision services, value guidance, and ecosystem operation.

[0023] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0024] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0025] Figure 1 A schematic diagram of the behavioral data acquisition prompt provided by the present invention is shown;

[0026] Figure 2 A flowchart of a first embodiment of the user profile determination method provided by the present invention is shown;

[0027] Figure 3 A flowchart of a second embodiment of the user profile determination method provided by the present invention is shown;

[0028] Figure 4 This illustration shows a data flow diagram provided by the present invention;

[0029] Figure 5 A flowchart of a third embodiment of the user profile determination method provided by the present invention is shown;

[0030] Figure 6 This illustration shows a data flow diagram provided by the present invention;

[0031] Figure 7 This illustration shows a data flow diagram provided by the present invention;

[0032] Figure 8A flowchart of a fourth embodiment of the user profile determination method provided by the present invention is shown;

[0033] Figure 9 A flowchart of a fifth embodiment of the user profile determination method provided by the present invention is shown;

[0034] Figure 10 This illustration shows a data flow diagram provided by the present invention;

[0035] Figure 11 A schematic diagram of a system architecture provided by the present invention is shown;

[0036] Figure 12 A schematic diagram of an embodiment of the user profile determination device provided by the present invention is shown;

[0037] Figure 13 A schematic diagram of an embodiment of the electronic device provided by the present invention is shown. Detailed Implementation

[0038] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0039] Addressing the technical issues of user profiling in related technologies that focus solely on short-term, immediate value and rely on explicit behavioral correlation analysis, making it difficult to uncover implicit user intentions and resulting in a one-sided profile, the inventors recognized that traditional modeling approaches based solely on surface-level behavioral statistics cannot reach users' true potential needs. Furthermore, the one-sidedness of the profiles makes it difficult to support long-term service decisions. Therefore, the inventors proposed starting from the intrinsic characteristics of user intentions, refining the representation of their ambiguous and ever-changing potential intentions, and simultaneously establishing a connection between micro-level behavioral characteristics and long-term macro-level value. This organically combines deep intention perception with macro-level value measurement, ultimately enabling the constructed user profiles to comprehensively consider users' implicit real needs and long-term value attributes, effectively improving the completeness and accuracy of the profiles, and providing reliable support for precise services with a long-term value orientation.

[0040] The user-related data involved in this invention, such as user behavior data, are all obtained after obtaining the user's permission or consent; that is, when this invention is applied to a specific product or technology, user permission is required to obtain and process the relevant data, and the processing of the relevant data must comply with the relevant laws, regulations and regulatory standards of the relevant countries and regions.

[0041] For example, Figure 1 The diagram illustrates the behavioral data acquisition prompts provided by the present invention, such as... Figure 1As shown, when it is necessary to obtain user behavior data, a behavior data acquisition prompt can be displayed on the vehicle's screen. After receiving confirmation from the user regarding the behavior data acquisition prompt, the controller can acquire the user's behavior data.

[0042] The execution subject of this invention can be an electronic device with processing capabilities, such as an in-vehicle system or a mobile terminal, but this invention does not limit the scope of the invention.

[0043] Figure 2 A flowchart illustrating a first embodiment of the user profile determination method provided by the present invention is shown. Figure 2 As shown, the method includes the following steps:

[0044] Step 110: Obtain the target user's behavior data for the current time period.

[0045] For example, the target user refers to a single user subject for whom a precise cognitive profile needs to be constructed. The current time period is a continuous statistical time window divided according to business needs, which can be defined by minute, hour, or calendar day. This embodiment of the invention does not limit the duration of the current time period.

[0046] Behavioral data includes multimodal interaction data, content consumption data, value conversion data, and ecosystem participation data generated by target users during that period. For example, it can include explicit behaviors such as page browsing, keyword search, content clicks, service subscriptions, community interaction, and sharing, as well as implicit trajectory features such as the time of occurrence, frequency, and duration of the behavior.

[0047] In one example, electronic devices can collect target user behavior data in real time through client-side front-end embeddings, or they can pull user behavior data for the current time period from a server-side behavior log library.

[0048] Step 120: Perform intent modeling and prediction processing on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period.

[0049] For example, intent modeling is a process that models user intent as a linear superposition of multiple independent intent ground states in a Hilbert space, and uses behavioral data to drive evolutionary operators to dynamically infer intent states. Predictive processing is a process that combines behavioral features with historical intent patterns to quantitatively fit fuzzy and potential user intents.

[0050] A micro-profile represents the weight distribution of a target user's intent components in the current time period. In other words, a micro-profile can be represented as a set of fine-grained features that focus on the user's underlying immediate needs and characterize the proportion of implicit intents. Here, intent components are fine-grained breakdowns of user needs, such as consumption intent, service usage intent, community participation intent, and content browsing intent. This embodiment of the invention does not limit the dimensions for dividing intent components. The weight distribution represents the activation intensity and dominance of each intent component through numerical methods.

[0051] In one example, an electronic device can map user behavior data to an initial quantum state in Hilbert space, construct a time evolution operator based on the temporal changes of behavior, perform evolution calculations on the previous intention superposition state to obtain the current intention superposition state, and then complete the collapse through the observation operator corresponding to the behavior data to obtain the basic intention state. The basic intention state can represent the probability distribution of the collapse to each intention ground state. The intention ground state is used to represent the potential interests of the target user, such as shopping, information, entertainment, etc. Subsequently, a feedforward neural network or temporal prediction model is used to predict the basic intention state and behavior data to obtain a microscopic profile including the probability distribution of each intention component.

[0052] In another example, after obtaining the basic intent state through intent modeling, the electronic device can process the basic intent state and the behavioral data of the current time period in parallel. On the one hand, the basic intent state and behavioral data are predicted through the aforementioned feedforward neural network or time-series prediction model to obtain each predicted intent and its confidence level. On the other hand, the basic intent state and behavioral data of the current time period are asserted in combination with preset business rules through first-order logical assertions to obtain the logical assertion results of each predicted intent. Then, the confidence levels of each predicted intent and the corresponding logical assertion results are aligned and arbitrated to obtain a micro-profile.

[0053] Step 130: Based on the preset macro value prediction model and the micro profile of the current period, determine the macro value indicators of the target user.

[0054] For example, macro-value metrics are long-term value quantification results that differ from short-term clicks and conversions. These metrics may include at least one dimension such as Lifetime Value (LTV), platform ecosystem contribution value, long-term retention value, and repurchase potential value. The macro-value prediction model is configured to learn a non-linear mapping from micro-level intent components to macro-value metrics. This macro-value prediction model can be used to achieve scale transformation from micro-level fine-grained features to macro-value dimensions. It should be noted that this embodiment of the invention does not limit the type of macro-value prediction model; for example, it can be a machine learning model such as a deep learning model.

[0055] For example, electronic devices can leverage graph neural networks and pre-defined value contribution maps to analyze users' connections and influences within social networks, quantifying the resulting ecological association characteristics. Based on a pre-defined intelligent agent simulation environment, they can simulate future user behavior paths according to historical micro-intentions and ecological association characteristics to predict macro-value indicators such as lifetime value, platform ecosystem contribution value, long-term retention value, and repurchase potential value. Furthermore, using historical user micro-intention data as input and simulation-derived macro-value indicator labels as supervisory signals, a loss function can be used to optimize the mapping relationship and train the macro-value prediction model.

[0056] Nonlinear mapping is a mathematical expression that uses multi-layer network transformations or kernel function fitting to characterize the complex nonlinear relationship between micro-level intentions and macro-level value indicators. This nonlinear mapping can, for example, be represented as a scaling function. Where M is a macroeconomic value indicator. It is an adjustable parameter, and m is a micro-feature vector (containing the weight distribution of each intention component in the micro-intention).

[0057] In one example, an electronic device can input a micro-profile into a trained deep neural network model (macro-value prediction model). The model outputs macro-value indicators such as the user's lifetime value score and ecological contribution level through multi-layer fully connected transformations and activation function mappings.

[0058] Step 140: Based on the micro-profile and macro-value indicators of the current time period, obtain the user profile of the target user.

[0059] For example, a user profile is a comprehensive cognitive result that integrates micro-level immediate intent characteristics with macro-level long-term value attributes. It includes both fine-grained characterization of users' potential needs and quantitative assessment of users' long-term value, and is structured data that comprehensively represents users' core characteristics.

[0060] In one example, electronic devices can directly concatenate and fuse the intent weight vector of the micro-profile with the macro-value indicator vector, and then eliminate the difference in dimensions through standardization to form a high-dimensional comprehensive user profile. Alternatively, an attention-weighted fusion method can be used, assigning different weights to micro-intent features and macro-value indicators according to the business scenario, and generating a structured user profile adapted to the value operation scenario after weighted summation.

[0061] This approach breaks through the limitations of static features and single time scale in traditional user profiling, constructing a user cognitive profiling system that integrates instantaneous micro-behavior and long-term macro-value. This system establishes two parallel profiling architectures: micro-behavioral particles and macro-value entities. The micro-behavioral particles correspond to the quantum intention state that changes dynamically at the second and minute levels, while the macro-value entities correspond to macro-value indicators such as long-term value and ecosystem contribution at the quarterly and annual levels. It can accurately reflect the dynamic changes of users' intentions at the second level and effectively assess the dynamic panoramic view of users' quarterly and even annual value potential, realizing cross-scale measurement and prediction of user value.

[0062] In this embodiment, after obtaining the target user's behavioral data for the current time period, a quantum intent modeling combined with time-series prediction is used to refine and calculate the user's implicit intent to obtain a micro-profile. Based on a pre-trained nonlinear macro-value prediction model, the micro-intent features are converted into long-term macro-value indicators. Then, by integrating the micro-intent and macro-value indicators, a comprehensive user profile is finally formed. This method achieves accurate quantitative characterization of the user's fuzzy implicit intent from the bottom layer, and introduces a long-term macro-value measurement dimension. It breaks through the one-sidedness of traditional user profiles that only focus on explicit behavior and short-term value, and greatly improves the comprehensiveness and accuracy of user profiles. This allows the profile to truly reflect the user's deep needs and long-term value attributes, providing more reliable data support and decision-making basis for subsequent precise services, value guidance, and ecosystem operation based on user cognition.

[0063] Figure 3 A flowchart illustrating a second embodiment of the user profile determination method provided by the present invention is shown. Figure 3 As shown, the method includes the following steps:

[0064] Step 210: Obtain the target user's behavior data for the current time period.

[0065] It should be noted that this step is similar to step 110 mentioned above, and will not be repeated here.

[0066] Step 220: Perform intent modeling on the behavioral data of the current time period to obtain intent state information.

[0067] For example, intent modeling is based on the principle of quantum superposition, which models user intent as a linear superposition of multiple independent intent ground states in Hilbert space. By driving time evolution operators with behavioral data, the intent state is dynamically deduced. At the same time, observation operators that match behavioral features are used to collapse the evolved intent superposition state to obtain quantifiable intent state results.

[0068] The intent base state represents the user's potential interests. In other words, the intent base state is a fine-grained breakdown of the user's potential interests, which may include subcategories such as shopping, entertainment, membership subscription, community creation, content consumption, and service consultation. This embodiment of the invention does not limit the dimensions of intent base state division.

[0069] Intent state information represents the weight distribution of at least one intent base state, wherein the weight distribution is a numerical representation of the activation intensity and relative proportion of each intent base state in the current time period; intent state information is a preliminary quantitative representation of the user's current potential intent, which can provide a basis for subsequent intent prediction and assertion.

[0070] In one example, an electronic device can map the feature vector of the behavior data of the current time period to the initial quantum state of Hilbert space, construct a time evolution operator based on the temporal changes of behavior, perform evolution calculation on the intention superposition state of the previous time period to obtain the intention superposition state of the current time period; then select an observation operator that matches the behavior type (such as the content consumption intention observation operator corresponding to browsing behavior), perform collapse processing on the intention superposition state, and output the weight distribution of each intention ground state, that is, the intention state information.

[0071] Step 230: Input the intent state information and the behavior data of the current time period into the preset neural network model to obtain the intent prediction information output by the neural network model.

[0072] For example, the preset neural network model is a deep learning model trained for fine-grained prediction of user intent. It uses historical behavior data and corresponding historical intent state information as input, and manually labeled real intents as tags to train the initial neural network model, resulting in the preset neural network model. This model uses intent state information and current time-segment behavior data as joint input, and predicts the intent component and its corresponding weights as output. It should be noted that this embodiment of the invention does not limit the type of neural network model; for example, it could be a Transformer model.

[0073] The intent prediction information represents the predicted intent components and the initial weight distribution. The predicted intent components are fine-grained demand units obtained by further decomposing the intent base state. The same intent base state can correspond to one or more predicted intent components, with a parent-child containment and refinement relationship. For example, if the intent base state is "buying a car," the predicted intent components can be refined to "SUV model consultation," "viewing new energy vehicle configurations," etc. The initial weight distribution is the model's initial quantification of the activation intensity of each predicted intent component.

[0074] In one example, Figure 4 The illustration shows a data flow diagram provided by the present invention, such as... Figure 4As shown, the electronic device is equipped with a neural network channel including a neural network model. The Transformer model can be used as the preset neural network model. The weight vector of the intent state information and the feature vector of the behavior data in the current time period are concatenated and input into the neural network channel. Feature extraction and intent pattern recognition are performed through multiple fully connected layers and activation functions, and the predicted intent components and the corresponding initial weight distribution are output.

[0075] Step 240: Determine the intent assertion information based on the intent state information, the behavior data of the current time period, and the preset business rule information.

[0076] For example, the business rule information is a set of structured rules built on domain common sense, behavioral constraints, and prior knowledge of intent association. For example, it may include behavior-intent association rules (such as "if a user browses the car model page for ≥20 minutes, then there is a fine-grained intent to inquire about purchasing a car"), intent validity verification rules (such as "if a user has no community interaction behavior, then all community-related fine-grained intents are invalid"), intent priority rules (such as "intent rules derived from paid behavior have higher priority than rules derived from browsing behavior"), abnormal intent filtering rules, etc. The embodiments of the present invention do not limit the content of the business rules.

[0077] Intent assertion information represents the assertion results of the predicted intent components. In other words, intent assertion information is a fine-grained intent component judgment result obtained through rule matching, with the same dimension as the intent prediction information. It includes the validity identifier (valid / invalid / suspicious) of each intent component, logical confidence (confidence score calculated based on rule matching degree, in the range of 0-1), rule priority level, and specific rule ID supporting the inference. It is used to judge the rationality and authenticity of the predicted intent components and to clarify whether each predicted intent component conforms to business logic and user behavior patterns.

[0078] In one example, refer to Figure 4 As shown, the electronic device is equipped with a symbolic reasoning channel running parallel to the neural network channel. The electronic device can input intent state information and current-time behavioral data into the aforementioned neural network channel, while simultaneously inputting them into this symbolic reasoning channel. Based on the core features of the intent state information and the current-time behavioral data, and combined with preset business rules, a first-order logical predicate deduction is performed through a logical reasoning engine to independently generate fine-grained intent component assertion results of the same dimension as the neural network channel output. Here, the first-order logical predicate is the basic expression form of symbolic reasoning, used to transform behavioral data and intent state information into inferable logical facts, such as predicate expressions like "behavior type (user ID, browsing)" and "intent ground state weight (user ID, car purchase, 0.7)".

[0079] Specifically, electronic devices can transform core features from the current period's behavioral data (such as behavior type = browsing, behavior object = new energy vehicle, behavior duration = 35 minutes, behavior frequency = 5 times) and core content from intent state information (such as intent base state = car purchase, intent base state weight = 0.75) into standardized first-order logical predicate facts, such as "behavior (user A, browsing, new energy vehicle, 35 minutes, 5 times)" and "intent base state (user A, car purchase, 0.75)". Then, the rule engine is invoked, pre-defined business rule information is loaded, and the first-order logical predicate facts are matched one by one with each rule in the business rule information. For example, if the rule "if the intent base state is car purchase" is matched... If a user browses a car model page for ≥20 minutes and ≥3 times, and the car model has a weight ≥0.6, then the fine-grained intent component 'New Energy Vehicle Inquiry' is considered valid, with a rule priority of Level 1 and a logical confidence level of 0.8. Simultaneously, the rule 'If the user does not engage in any car purchase-related payment behavior, then the fine-grained intent component 'Place a Car Order' is considered invalid, with a rule priority of Level 2 and a logical confidence level of 0.9' is matched. Through inductive and deductive reasoning, the judgment results of all fine-grained intent components are derived one by one. Each intent component is labeled with its validity, logical confidence level, supporting rule ID, and priority. Finally, all judgment results are integrated to form structured and interpretable intent assertion information.

[0080] Step 250: Perform semantic alignment and conflict arbitration on the intent prediction information and intent assertion information to obtain the micro-profile of the target user in the current time period.

[0081] For example, semantic alignment processing involves matching and unifying the predicted intent component in the intent prediction information with the assertion result in the intent assertion information at the semantic level, eliminating the differences between the two in intent expression and dimensional division, and ensuring information consistency.

[0082] Conflict arbitration is a process of selection and correction based on pre-set arbitration logic (such as business priority and data credibility) when there is a contradiction between the intention prediction information and the intention assertion information (such as the model predicting that a certain intention component is valid, but the assertion result is invalid).

[0083] In one example, refer to Figure 4As shown, the electronic device can also be equipped with a semantic alignment and conflict resolution center. It can first calculate the semantic similarity between the predicted intent component in the intent prediction information and the predicted intent component in the intent assertion information through a multi-head attention mechanism, and map the neural-side predicted intent component and the symbol-side assertion intent component to the same semantic space to complete the dimension alignment. After alignment, the intent component conclusions under the same semantic dimension are compared one by one to determine whether there is a conflict. If the two results are consistent (no conflict), the weights are directly merged and normalized. If there is a conflict, a multi-level conflict resolution strategy is initiated. First, the confidence weight of the neural output and the priority of the symbol rule are compared. If it is still impossible to determine, the context re-evaluation is carried out in combination with the current time period behavior context and intent state information. Then, the source conclusion with higher confidence after re-evaluation is adopted. If there is still an unresolved conflict after the above process, the manual review flag is automatically triggered and the corresponding intent component is marked. Finally, after all processing, a stable, consistent and interpretable micro-profile of the target user in the current time period is output.

[0084] refer to Figure 4 As shown, a dual-channel parallel architecture containing neural network channels and symbolic reasoning channels is constructed. Through semantic alignment and conflict resolution centers, the semantic similarity between neural outputs and symbolic conclusions is calculated using an attention mechanism. When results conflict, dynamic arbitration is performed based on preset confidence weights and rule priorities to output a final consistent and interpretable reasoning result.

[0085] Optionally, the electronic device can optimize the system parameters of the neural network channel and the symbolic reasoning channel based on the micro-profile, intent state information and behavioral data of the current time period. Through the feedback learning mechanism, the system can have the ability to continuously evolve, thereby enhancing its feasibility and dynamism.

[0086] Step 260: Based on the preset macro value prediction model and the micro profile of the current period, determine the macro value indicators of the target user.

[0087] It should be noted that this step is similar to step 130 mentioned above, and will not be repeated here.

[0088] Step 270: Based on the micro-profile and macro-value indicators of the current time period, obtain the user profile of the target user.

[0089] It should be noted that this step is similar to step 140 mentioned above, and will not be repeated here.

[0090] Step 280: Based on the user profile of the target user, perform matching processing in the preset strategy library to determine the personalized intervention strategy for the target user.

[0091] For example, the pre-defined strategy library is a set of structured intervention strategies based on business operation goals and user value enhancement needs. It may include various types of intervention strategies such as personalized pricing, incentive programs, content guidance, community operations, retention and activation, and ecosystem contribution guidance. Each intervention strategy corresponds to specific applicable user profile characteristics. Matching is the process of selecting intervention strategies from the strategy library that match user needs and guide users towards higher-value paths (improving macro-value indicators) based on the micro-intent distribution and macro-value indicators of the target user profile. Personalized intervention strategies are customized specifically for target users to improve their macro-value indicators.

[0092] In one example, electronic devices can use a feature matching algorithm to calculate the similarity between the feature vector of the user profile and the applicable feature vector of each intervention strategy in the strategy library, and select the intervention strategy with the highest similarity as the personalized intervention strategy.

[0093] Step 290: Implement personalized intervention strategies.

[0094] For example, implementing personalized intervention strategies is the process of applying a defined, customized intervention plan to user interaction scenarios. Appropriate outreach methods and execution timing can be selected based on user profile characteristics and scenario features to ensure the effectiveness of the intervention strategy and user experience. For instance, if a personalized intervention strategy is a membership benefit incentive, the weight of membership intent in the user's micro-profile can be used to reach the user through pop-ups, push notifications, etc., when the user browses membership-related content, and to perform benefit distribution and guidance operations. Some pre-planned strategies are community creation incentives; after a user publishes content, operations such as points rewards and traffic support can be implemented based on their ecosystem contribution value in the macro-value indicators.

[0095] In this embodiment, behavioral data of the target user in the current time period is acquired, and intent state information is obtained through intent modeling. Intent prediction information is output by combining behavioral data with a neural network model. Intent assertion information is determined based on intent state, behavioral data, and business rules. A micro-profile is obtained through semantic alignment and conflict arbitration. Then, a macro-value prediction model derives macro-value indicators from the micro-profile. The two are integrated to form a complete user profile. Based on this user profile, personalized intervention strategies are matched and executed from a preset strategy library. This method relies on quantum modeling to achieve a refined and precise characterization of users' implicit intent, improves the interpretability of intent reasoning through neural symbol fusion, constructs a comprehensive user profile by combining micro-intent and macro-value, and proactively guides users to enhance their value through personalized intervention strategies. This effectively overcomes the limitations of traditional methods, such as difficulty in characterizing users' implicit intent, lack of long-term value measurement, and lack of targeted intervention strategies. It significantly improves the comprehensiveness and accuracy of user profiles, enhances the adaptability and effectiveness of intervention strategies, and can deeply understand users' real needs and guide users to migrate towards high-value paths, providing reliable technical support for the platform's long-term value operation and precise service output.

[0096] Figure 5 The flowchart illustrates a third embodiment of the user profile determination method provided by the present invention. Figure 6 The illustration shows a data flow diagram provided by the present invention, such as... Figure 5 and Figure 6 As shown, step 220 above may include the following sub-steps:

[0097] Step 310: Determine the time evolution operator based on the behavioral data of the current time period.

[0098] For example, the time evolution operator is a mathematical operator used to describe the dynamic changes of the superposition state of user intent as the behavior changes over time.

[0099] In one example, to ensure that the superposition state of user intent can dynamically change according to real-time behavior according to quantum mechanical rules, the time evolution operator was not manually calculated using a fixed formula. Instead, a differentiable neural network module was designed to automatically generate time evolution operators. Specifically, this differentiable neural network module includes a feature extraction and control network, an equivalent Hamiltonian construction unit, and a differentiable unitary transform network. It inputs the multimodal behavioral data of the current time period into the feature extraction and control network for feature extraction and encoding, and outputs dynamic parameters to characterize the cognitive environment. Based on the equivalent Hamiltonian construction unit, an equivalent Hamiltonian representing the user's current cognitive environment is constructed according to this dynamic parameter. Based on a differentiable unitary transform network, according to a preset discrete time step... And the equivalent Hamiltonian, through matrix exponentiation according to the formula An approximate solution is performed to obtain the time evolution operator. Meanwhile, the evolution operator is constrained to satisfy the unitary transformation property, ensuring that the evolution of the intended state conforms to the rules of quantum mechanical coherence and probability preservation; the unit has differentiable properties and can be used in conjunction with gradient updates to complete parameter optimization.

[0100] The entire computation process, from data input to operator output, is encapsulated into a differentiable neural network module. This design allows the system to learn the mapping relationship between behavioral data and intention state evolution from end to end through data training, ensuring that the evolution of the intention superposition state strictly follows the coherence and probability conservation rules of quantum mechanics, and ultimately accurately simulates the dynamic changes in the uncertainty of user intention as behavior occurs.

[0101] Step 320: Perform temporal evolution processing on the intent superposition state information of the target user in the previous time period based on the time evolution operator to obtain the intent superposition state information in the current time period.

[0102] For example, the intent superposition state information of the previous time period is the intent quantum state of the target user in the previous time period, which is characterized by multiple complex probability amplitudes.

[0103] Temporal evolution processing is a dynamic evolution process that relies on time evolution operators to perform unitary transformation on the intention superposition state, adjusting the magnitude and phase of the probability amplitude of each intention ground state.

[0104] The intention ground state is an orthogonal basis vector in Hilbert space representing a class of user intentions, which may include, for example, […]. , , Etc. Intention superposition state information is a linear superposition representation of at least one intention ground state, for example, it can be represented as... ,in, , , This represents the complex probability amplitude of the proportion of each intention's ground state. Understandably, based on the principle of superposition and the concept of entanglement in quantum mechanics, user intentions can be represented as state vectors in Hilbert space. For example, by constructing a Hilbert space model of user intentions, multimodal behavioral data can be mapped to ground states in this space through feature encoding. Thus, user intentions can be represented as a linear superposition of multiple ground states. It should be noted that the initial probability amplitude (coefficient) of the intention superposition state can be initialized based on the user's initial behavioral characteristics or prior distribution. For example, based on user registration information, historical behavior statistics, etc., different intention ground states can be assigned an initial amplitude and phase.

[0105] Understandably, the evolutionary process follows the principles of coherence and probability preservation in quantum mechanics. It strengthens the probability amplitude of the ground state with associated intentions and weakens the probability amplitude of the ground state with irrelevant intentions based on behavioral data, while adjusting the phase to incorporate contextual information.

[0106] In one example, the target user's intent from the previous time period is overlaid with state information. The unitary transformation operation is performed with the time evolution operator. During the evolution process, based on the directionality of the behavioral data in the current time period, the modulus of the complex probability amplitude corresponding to the ground state of the associated intention is increased and its phase is rotated, while the complex probability amplitude of the ground state of the unassociated intention is simultaneously decreased. Finally, the superposition state information of the current time period intention, which preserves quantum coherence and normalizes the probability, is obtained. ).

[0107] Specifically, when an electronic device receives new multimodal behavioral data for the current time period, such as a user's single click, generated text content, or viewed image information, it encodes this multimodal behavioral data into a set of operations acting on the intention state Hilbert space. This set of operations serves as the influence factor (equivalent Hamiltonian) of the time evolution operator. Based on the actual content of the behavioral data, the magnitude and phase of the complex probability amplitude corresponding to each intention ground state are adjusted. For example, if the current multimodal behavioral data strongly points to a certain intention ground state, such as when a user frequently browses a product details page and corresponds to the |Shopping> intention ground state, the magnitude of the complex probability amplitude corresponding to that intention ground state will increase accordingly, and the phase will rotate to incorporate new contextual information. For other intention ground states that are unrelated to the current multimodal behavioral data, their corresponding complex probability amplitudes will decrease accordingly, thereby realizing the dynamic evolution of the intention superposition state.

[0108] Step 330: Based on the observation operator corresponding to the behavior data of the current time period, perform collapse processing on the intent superposition state information of the current time period to obtain intent state information.

[0109] For example, when a specific inference needs to be made (such as predicting the user's next action), an observation is performed. This process corresponds to defining an observation operator that is specific to a user interaction behavior (such as clicking the "buy" button). The observation operator is a quantum measurement operator that matches the type of user interaction behavior in the behavioral data of the current time period. It is used to trigger wave function collapse and thus achieve observation collapse of the intention superposition state.

[0110] Collapse is a quantum measurement process that converges an uncertain superposition of intentions into a probability distribution of the ground states of those intentions. The probability calculation follows the rules of quantum mechanics, collapsing to a specific ground state of the intention. probability Given the corresponding complex probability amplitude c_i (i.e. , , The square of the modulus of (etc.) is determined, that is , or expressed as It quantifies the probability distribution from an uncertain superposition state to a definite intention.

[0111] The complex probability amplitudes are not obtained through classical methods such as weighted averaging, but are generated by a quantum-style encoder. This encoder takes behavioral data as input, maps it to Hilbert space through a pre-defined deep neural network, and outputs a normalized complex state vector. The projection components of the state vector onto each intention ground state are the complex probability amplitudes of each intention ground state. This generation method simulates the "state preparation" process of a quantum system. The generated intention superposition states themselves can encode complex, non-classical correlations and interference relationships between multimodal behavioral data, laying a truly quantum mechanical mathematical foundation for subsequent quantum evolution and observation processing.

[0112] Intent state information is a set of ground state probability distributions of each intent obtained after collapse processing, and is a quantifiable representation of user intent.

[0113] In one example, a dedicated observation operator is first matched based on the user interaction type corresponding to the behavior data of the current time period. This observation operator is then applied to the intent superposition information of the current time period to perform quantum measurement, triggering superposition collapse. Finally, the collapse probability is obtained by calculating the square of the modulus of the probability amplitude corresponding to each intent ground state through the complex probability amplitude generated by the quantum style encoder. After normalizing and regularizing the probability results, structured intention state information containing the ground state probability distribution of each intention is formed.

[0114] The above steps fundamentally reshape user intent using a quantum mechanical paradigm. Specifically, by mapping multimodal user behavior data to a Hilbert space, each potential point of interest is defined as the ground state of intent, while the user's overall intent is represented as a linear superposition of these ground states. When a user interacts (i.e., "observes"), the superposition collapses with a specific probability amplitude into a particular ground state. Simultaneously, quantum probability precisely characterizes the dynamic process of intent evolving from uncertainty to certainty, providing a deep cognitive foundation that effectively handles ambiguity and superposition, far exceeding traditional probabilistic models.

[0115] In this embodiment, a time evolution operator is obtained by feature processing and Hamiltonian construction of multimodal behavior data in the current time period. This time evolution operator is used to perform quantum mechanical coherent probability-preserving evolution on the superposition state of the user's intention in the previous time period. The modulus and phase of the complex probability amplitude of each intention ground state are adjusted to obtain the superposition state in the current time period. Then, measurement collapse is performed by the observation operator matched with the interaction behavior. The intention state information is obtained by calculating the square of the modulus of the probability amplitude. The preparation, evolution and measurement process of the intention state is constructed in strict accordance with the principles of quantum mechanics. By encoding the complex correlation of multimodal data with complex probability amplitude, the superposition, uncertainty and dynamic change characteristics of the user's intention are accurately simulated. The classical weighted average coefficient generation method is abandoned, which greatly improves the refinement and scientificity of the user's intention state characterization and provides basic data that conforms to the evolution law of real intention for intention reasoning.

[0116] In conclusion, Figure 7 The illustration shows a data flow diagram provided by the present invention, such as... Figure 7 As shown, when a user generates real-time behavioral data (such as triggering events like searching for the range of new energy vehicles), the electronic device first inputs the real-time behavioral data into the quantum sensing module. The quantum sensing module calculates the user's current time period's intent superposition state information based on the input behavioral data. According to the complex probability amplitude corresponding to each intent ground state in the superposition state, the intent state information is obtained. Then, through neural network channels and symbolic reasoning channels, prediction and logical reasoning are performed to obtain the micro-profile of the current time period, completing the update of the micro-profile state in the micro-profile library, forming a micro-profile that represents the probability distribution of the current user's intent.

[0117] After completing the state update, the micro-profile library passes the current micro-feature vector, containing the weight distribution of each intent component, to the scaling transformation function module. The scaling transformation function module uses this micro-feature vector, based on a pre-defined macro-value prediction model (scaling transformation function), to predict macro-value indicators (macro-profiles) such as the user's lifetime value and platform ecosystem contribution level. Based on the current time-period micro-profile and macro-value indicators, a cross-scale user profile of the target user is obtained and stored in the user profile library. During training, the macro-value prediction model can predict macro-value indicators corresponding to the user's historical behavior data based on the macro-profile predictor and value graph engine, obtaining sample macro-value indicators for model training. Simultaneously, after determining the macro-value indicators based on the current time-period micro-profile, the target user's value contribution graph can be updated synchronously.

[0118] Furthermore, personalized intervention strategies can be determined and implemented based on cross-scale user profiles of target users, and then real-time behavioral data of users can be collected to update user profiles in the future.

[0119] The inventors also discovered that traditional micro-intention modeling is mostly driven by users' instantaneous multimodal behavior, and the update of intention weights is only based on the short-term matching degree of the current interaction, which has significant short-sighted optimization defects: this type of modeling only focuses on the user's current behavioral tendencies and fails to establish an effective connection between the evolution of micro-intention and macro-value goals such as the user's long-term life cycle value and contribution to the ecosystem network. This can easily lead to service strategies that excessively cater to users' immediate preferences, fall into an "information cocoon" or a short-term traffic trap, and ultimately overdraw on the user's long-term value and deviate from the core goal of personalized services. Therefore, this invention proposes to use macro-value indicators as the core optimization objective to dynamically reshape the probability distribution and weight ratio of micro-intents. This ensures that the evolution of micro-intents is no longer driven solely by instantaneous behavior, but is always dynamically calibrated with the goal of maximizing long-term value. This approach retains the ability to accurately perceive users' real-time intentions at the micro level, while also achieving bidirectional collaborative optimization between micro-behavioral particles and macro-value entities through the reverse constraints and guidance of macro-values. This fundamentally solves the short-sightedness problem of traditional intent modeling, resulting in a more accurate and comprehensive user profile. Consequently, it ensures that subsequent personalized service strategies always revolve around maximizing long-term user value and the overall value of the ecosystem.

[0120] In some embodiments, before performing intent modeling and prediction processing on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period, the electronic device may generate a value gradient signal based on the micro-profile of the target user in the previous time period when preset conditions are met.

[0121] Among them, the value gradient signal is used to adjust the weight distribution of each intention component in the micro-profile during the current time period; the preset conditions include reaching a preset period or triggering a preset key node.

[0122] For example, the preset conditions are the criteria for determining the generation of value gradient signals, including two types: reaching a preset time statistical period and key nodes triggered by user behavior or business scenarios. The preset period is a fixed time window divided according to business needs, and key nodes include preset value-related triggering events such as users browsing core services, initiating business inquiries, and completing interactive conversions.

[0123] The value gradient signal is a guiding control signal generated based on the relationship between a user's historical intent and value. It is used to apply directional adjustments to the weight distribution of each intent component in the current micro-profile during subsequent quantum evolution and intent weight calculation, focusing on strengthening intent components that are highly correlated with macro-value and weakening intent components with low value correlation.

[0124] The micro-profile of the previous period is a set of weight distributions of intent components obtained after semantic alignment and conflict arbitration in the previous period.

[0125] In one example, the electronic device determines whether the current time period threshold is met or whether the target user has triggered a value-related key business node. If either preset condition is met, the device retrieves the micro-profile of the target user from the previous time period. Based on the intention components in the micro-profile from the previous time period, and combined with the macro-value index association rules, the device calculates the contribution of different intention components to the macro-value index. Based on the magnitude of the contribution, a value gradient signal containing gain weights and inhibition weights is generated. This signal is then fed into the subsequent intention modeling and prediction processing flow to achieve targeted control during the intention component weight fitting stage of the current time period.

[0126] In another example, Figure 8 A flowchart of a fourth embodiment of the user profile determination method provided by the present invention is shown, as follows: Figure 8 As shown, this step may include the following sub-steps:

[0127] Step 410: Based on the preset intelligent agent, perform simulation prediction processing on the micro profile of the previous period to obtain the first future macro value index corresponding to the micro profile of the previous period.

[0128] For example, the preset intelligent agent is a simulation reasoning intelligent agent built based on the macro value prediction model and user behavior evolution rules, which has the ability to extrapolate the future macro value of user intention characteristics.

[0129] Simulation prediction processing uses the micro-profile of the previous time period as input and performs multi-step forward extrapolation along the logic of user behavior and value evolution to simulate the change process of future macro-value indicators under the intention distribution represented by the micro-profile of the previous time period. The first future macro-value indicator M_current is the quantitative result of macro-value indicators such as future long-term life cycle value, ecological contribution value, and retention potential obtained by the agent based on the micro-profile of the previous time period.

[0130] In one example, the original micro-profile from the previous period is input into a preset simulation agent. The agent performs multi-step future scenario simulation and deduction based on the weight distribution of each intent component in the profile, combined with the user's historical behavior patterns and macro-value mapping relationship. It outputs quantitative indicators such as user lifecycle value score and ecosystem contribution level for a corresponding period of time in the future, forming the first future macro-value indicator.

[0131] Specifically, based on a pre-defined graph convolutional network and a pre-defined value contribution map, the ecological association characteristics of the target user are determined; the micro-profile and ecological association characteristics of the previous time period are input into the agent to simulate the first behavioral data of the target user in the future time period under the intent represented by the micro-profile of the previous time period; based on the agent, intent modeling and prediction processing are performed on the first behavioral data of the future time period to obtain the first micro-profile of the future time period; based on the agent, the first future macro-value index is obtained according to the first micro-profile of the future time period and the macro-value prediction model.

[0132] The value contribution graph is a heterogeneous graph constructed based on the target user's historical behavior data, social relationship data, and content dissemination link data. For example, it can use users, content, social objects, and dissemination nodes as entities, and interactions, follows, reposts, likes, and behavioral associations as edges to build a heterogeneous graph structure, used to represent the user's value association and influence within the ecosystem. A graph convolutional network is used to aggregate neighborhood features and infer relationships within this value contribution graph, extracting structured features such as the user's influence, dissemination ability, and association value within the ecosystem. Ecosystem association features are comprehensive characteristics reflecting the target user's association strength, dissemination potential, social value, and group influence within the platform ecosystem.

[0133] The first behavioral data for the future time period is a sequence of simulated interactions, browsing, dissemination, and consumption behaviors generated by the agent based on the micro-profile and ecological relationship features of the previous time period. The first micro-profile for the future time period is derived internally by the agent based on the first behavioral data of the future time period obtained from simulation, and is used to represent the weight distribution of future intention components.

[0134] For example, electronic devices can pre-construct a value contribution graph containing various node and edge types from the target user's historical behavior data, social relationship data, and content propagation link data. This value contribution graph is then input into a pre-trained graph convolutional network. Through multi-layer graph convolution, node features and neighborhood relationships are aggregated and inferred, outputting an ecological association feature vector corresponding to the target user. The intent weight features of the micro-profile from the previous time period are concatenated and fused with the ecological association features, and input into the agent. The agent determines behavioral preferences based on the intent distribution of the micro-profile from the previous time period, and combines the ecological association features to constrain the behavioral range and propagation path, simulating and generating continuous and logically consistent user behavior time-series data for the future time period (the first behavioral data of the future time period). The agent performs quantum evolution calculations on the simulated first behavioral data of the future time period, constructs a corresponding time evolution operator, updates the historical intent superposition state, and then performs collapse processing using an observation operator matching the simulated behavior. After semantic alignment and conflict arbitration, the first micro-profile of the future time period is obtained. The intelligent agent inputs the generated first micro-profile of the future period into the pre-trained macro-value prediction model, and obtains quantitative indicators such as the user's future life cycle value, ecological contribution value, and retention potential through nonlinear mapping reasoning. These indicators are the corresponding first future macro-value indicators.

[0135] By constructing a pre-built user value contribution map and extracting ecological association features using graph convolutional networks, and combining the micro-profile of the previous period to drive the agent to complete future behavior simulation, the agent then obtains a future intention profile through a quantum evolution process within the agent. Finally, the first future macro-value index is output based on the macro-value prediction model. This approach integrates ecological association, intention evolution, and value prediction into the agent simulation chain, making the derivation of the first future macro-value index more consistent with real ecological constraints and the dynamic changes in intentions, thereby improving the accuracy and reliability of value gradient signal generation.

[0136] Step 420: Based on the intelligent agent, perform simulation prediction processing on the adjusted micro-profile to obtain the second future macro-value index corresponding to the adjusted micro-profile.

[0137] The adjusted micro-profile is obtained by adjusting the weight of the target intent component in the micro-profile of the previous period. For example, the intention weight of "learning content" in the micro-profile of the previous period is increased, and the intention weight of "low-quality entertainment content" is decreased to obtain the adjusted micro-profile.

[0138] For example, the target intent component is an intent unit selected from the micro-profile of the previous time period that has a potential correlation with the improvement of the macro-value indicator. The number of target intent components is not limited. The adjusted micro-profile is a comparison sample obtained by adjusting the weight of the target intent component based on the micro-profile of the previous time period, while keeping the weights of other intent components unchanged. The second future macro-value indicator M_simulated is the quantitative result of future macro-value obtained by the agent based on the simulation of the micro-profile with adjusted weights. It is used to compare with the first future macro-value indicator to reflect the degree of impact of the change in the weight of the intent component on the value.

[0139] In one example, the weights of the target intent component in the micro-profile of the previous time period are slightly increased or decreased while maintaining overall normalization to obtain the adjusted micro-profile. The adjusted micro-profile is then input into the same simulated agent and predicted using the same simulation logic and inference step size as the first future macro-value indicator to obtain the corresponding second future macro-value indicator.

[0140] Specifically, after obtaining the ecological association features, the adjusted micro-profile and ecological association features are input into the intelligent agent to simulate the second behavioral data of the target user in the future time period under the intent represented by the adjusted micro-profile. Based on the intelligent agent, the second behavioral data in the future time period is modeled and predicted to obtain the second micro-profile in the future time period. Based on the intelligent agent, the second future macro value index is obtained according to the second micro-profile in the future time period and the macro value prediction model.

[0141] The second behavioral data for the future time period consists of simulated interaction, browsing, dissemination, and consumption sequences generated by the agent based on the adjusted micro-profile and ecological association features within a future time window. The second micro-profile for the future time period is derived internally by the agent based on the simulated second behavioral data for the future time period and is used to characterize the weight distribution of future intention components.

[0142] For example, an electronic device concatenates and fuses the intent weight features of the adjusted micro-profile with the aforementioned ecological association features, and inputs this data into an intelligent agent. The agent determines behavioral preferences based on the intent distribution of the adjusted micro-profile, and constrains the scope and propagation path of behavior by combining ecological association features, simulating and generating continuous and logically consistent time-series data of user behavior in future time periods (second behavioral data for future time periods). The agent performs quantum evolution calculations on the simulated second behavioral data for future time periods, constructs corresponding time evolution operators, updates the superposition state of historical intents, and then performs collapse processing using observation operators matching the simulated behavior. After semantic alignment and conflict arbitration, the second micro-profile for future time periods is obtained. The agent inputs the generated second micro-profile for future time periods into a pre-trained macro-value prediction model, and obtains quantitative indicators such as the user's future lifecycle value, ecological contribution value, and retention potential through nonlinear mapping inference. These indicators are the corresponding second future macro-value indicators.

[0143] Optionally, a target intent component can be selected from multiple intent components contained in the micro-profile of the target user in the previous time period, based on a preset value contribution graph. The value contribution graph is a heterogeneous graph constructed based on historical behavior data, social relationship data, and content propagation link data. The graph records the user behavior nodes, social connection nodes, and content propagation nodes corresponding to each intent component, as well as the strength of the value association between them. The target intent component is the intent component in the micro-profile that has a strong correlation with high-value nodes in the value contribution graph and can have a significant impact on the user's macro-value indicators. For example, each intent component in the micro-profile of the previous time period can be mapped to the corresponding behavior nodes and propagation connection nodes in the value contribution graph. The centrality of the associated nodes, value transmission weight, and ecological influence range of each intent component in the graph can be calculated. Intent components with a correlation strength higher than a preset threshold or a high ranking in value transmission weight can be selected as the target intent components for subsequent weight adjustments.

[0144] Step 430: Generate a value gradient signal based on the first future macroeconomic value indicator and the second future macroeconomic value indicator.

[0145] For example, the numerical differences between the first and second future macroeconomic value indicators in dimensions such as long-term life cycle value and ecological contribution value, as well as the comprehensive value difference, are calculated respectively. The positive or negative value of the difference determines the effect of adjusting the target intention component weight on improving or inhibiting the macroeconomic value. The magnitude of the difference determines the intensity of regulation, forming value-oriented information that includes the direction of value optimization and the magnitude of regulation. Subsequently, this value-oriented information, along with the corresponding intention component adjustment identifier, is used as a monitoring signal or reward signal and input into the backpropagation process of the aforementioned scale transformation function f or a preset strategy network. Within the framework of differentiable computation, the gradient of the comprehensive value difference with respect to each intention component in the micro-intention feature vector m is solved by the chain rule of differentiation, thereby obtaining the optimized gradient of the weight of each intention component. Alternatively, in the policy network, with the maximization of the value difference as the optimization objective, the weight adjustment strategy of each intention component is generated through policy iteration and decision reasoning. This maps the optimization requirements at the macro-value level into weight adjustment rules that can be directly executed at the micro-intention dimension. Finally, the value gradient signal with clear direction and magnitude is integrated to guide the adjustment of the weight of the intention components during the evolution of the intention superposition state and the generation of the micro-profile in the current time period.

[0146] In this embodiment, a pre-set simulation agent is used to simulate and extrapolate the future macro-value of the micro-profile from the previous time period and the micro-profile after weight adjustment, respectively, to obtain two sets of comparative future macro-value indicators. Then, a value gradient signal is generated based on the difference between the two sets of indicators. This method quantifies the actual contribution of different intent components to the user's future macro-value indicators through simulation comparison, so that the value gradient signal has an objective basis for extrapolation. It can accurately guide the distribution of intent weights in a direction that is conducive to improving macro-value in the subsequent intent modeling process. This makes the micro-profile not only fit the user's real-time behavior, but also has a clear long-term value orientation, providing a more valuable and targeted intent basis for subsequent personalized intervention strategies.

[0147] Furthermore, after obtaining the value gradient signal, refer to Figure 7 As shown, the weights of each intent component in the micro-profile can be dynamically adjusted based on the value gradient signal during the subsequent evolution of the micro-profile based on real-time user behavior data. In one possible implementation, the electronic device can use the value gradient signal as the arbitration logic parameter in the conflict arbitration process of the aforementioned semantic alignment and conflict resolution center. That is, the intent prediction information and intent assertion information are first semantically aligned to obtain aligned intent prediction information and aligned intent assertion information; based on the value gradient signal in the preset arbitration logic parameter, the aligned intent prediction information and aligned intent assertion information are arbitrated to obtain the micro-profile of the target user in the current time period.

[0148] In this implementation, the value gradient signal can participate in the conflict arbitration process as an arbitration logic parameter. When there is a conflict between intention prediction information and intention assertion information, it provides a value-oriented basis for arbitration decisions from the perspective of long-term macro-value enhancement, i.e., it tends to select the intention component that can improve the macro-value indicator. At this point, the arbitration process, based on semantic alignment, combines multi-dimensional conditions such as the value gradient signal, confidence weight, and rule priority to select, adjust, or merge conflicting intention items, ultimately obtaining a logically consistent and value-oriented intention result.

[0149] Specifically, the intent components in the intent prediction information and intent assertion information are first semantically matched and dimensionally normalized to achieve semantic alignment and mark intent items with conflicting results. For intent items without conflict, weight fusion is directly performed. For intent items with conflict, the value gradient signal in the preset arbitration logic parameters is retrieved, and a comprehensive judgment is made in combination with the neural network prediction confidence, symbol rule priority, and contextual behavior information. Based on the direction of the value gradient signal, the intent components that are more conducive to improving future macro value are retained first and their weights are strengthened. Intent components that are contrary to the direction of value improvement are suppressed or eliminated. After arbitrating all conflicting items, the overall intent weights are normalized and regulated to obtain the micro profile of the target user in the current time period.

[0150] This approach introduces value gradient signals as logical parameters into the conflict arbitration process. When conflicts arise between intention prediction and intention assertion information, it not only relies on confidence levels or rule priorities for decision-making but also simultaneously incorporates long-term value orientation to complete intention selection and weight adjustment. This results in a micro-profile that is not only logically consistent and accurate in intention recognition but also naturally aligns with the user's macro-value enhancement goals. This method imbues the conflict resolution process with a clear value orientation, enabling the micro-profile to simultaneously meet the dual requirements of high-precision intention characterization and value optimization orientation. It provides a more value-oriented intention foundation for subsequent personalized intervention strategies, further enhancing the effectiveness of these strategies in improving macro-value indicators.

[0151] In another possible implementation, the electronic device can optimize the system parameters based on the value gradient signal to obtain optimized system parameters; then, based on the optimized system parameters, it can perform intent modeling and prediction processing on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period.

[0152] The system parameters include at least one of the following: quantum parameters in intent modeling, model parameters in the neural network model, and arbitration logic parameters in conflict arbitration. Quantum parameters include parameters related to quantum state evolution and collapse, such as time evolution operators, equivalent Hamiltonians, observation operators, and complex probability amplitude initialization rules. Neural network model parameters include learnable parameters used for intent prediction, such as Transformer feature extraction layer weights, fully connected layer parameters, and attention weight coefficients. Conflict arbitration logic parameters include parameters used for conflict resolution, such as confidence thresholds, rule priority weights, context re-evaluation coefficients, and conflict trade-off criteria.

[0153] In this implementation, the value gradient signal can be used to guide the system parameters to iterate in a direction more conducive to improving macro-value. System parameter optimization is a process of updating, correcting, and adapting relevant parameters based on the value gradient signal, aiming to ensure that the entire subsequent intent modeling process has value-oriented capabilities. The optimized system parameters are a set of parameters updated after incorporating value guidance, enabling the entire process of intent modeling, intent prediction, and conflict arbitration to serve the goal of improving macro-value.

[0154] In one example, the electronic device resolves the value gradient signal into the value gain direction and modulation intensity of the corresponding intention component, and performs targeted optimization on at least one type of system parameter: for quantum parameters, it strengthens the probability amplitude gain coefficient of the ground state of high-value associated intention based on the value gradient signal, and adjusts the phase and amplitude parameters of the corresponding intention component in the evolution operator; for neural network model parameters, it performs small-step gradient updates with the value gradient signal as an additional supervision signal, so that the model is more biased towards the intention corresponding to high macroscopic value indicators when predicting intention components and weights; for conflict arbitration logic parameters, it increases the arbitration weight of value-sensitive intention components based on the value gradient signal, and adjusts the conflict judgment threshold and selection rules. After completing all optimizations, the optimized system parameters are obtained. The optimized system parameters are then applied to the quantum evolution, observation processing, neural network intention prediction, and subsequent conflict arbitration process of the current time period behavioral data, and finally obtains a micro-profile of the current time period that combines intention recognition accuracy and value orientation.

[0155] By uniformly optimizing system parameters such as quantum evolution parameters, neural network model parameters, and conflict arbitration parameters through value gradient signals, the entire intent modeling chain has a clear long-term value orientation, tilting towards improving macro value. This ensures that the micro-profile accurately depicts the user's true intent and achieves a deep binding between intent representation and long-term value goals, effectively improving the value conversion efficiency of subsequent personalized intervention strategies and the improvement effect of macro value indicators.

[0156] In this embodiment, through the above-described method, based on the value gradient signal adjustment, the electronic device, in the process of generating the micro-profile for the current time period, will specifically enhance the complex probability amplitude of positive intent components that can continuously improve the user's long-term lifecycle value, ecological contribution value, and network effect value (such as "becoming a car club member," "participating in high-quality community creation," "actively spreading content," etc.) in the intent superposition state, so that these high-value-oriented intents will obtain a higher activation ratio in the micro-intent distribution. Conversely, for negative intents that are likely to lead to a decrease in user retention, a decrease in interaction quality, or even damage to the platform's long-term ecological value (such as "frequent page bounces," "low-quality shallow interactions," "meaningless violations," etc.), the probability amplitude of these intents will be suppressed and attenuated through the value gradient signal, reducing the influence ratio of such intents in the intent superposition state. Ultimately, the micro-intent for the current time period is obtained, which not only fits the user's real-time behavioral characteristics but also always anchors to the maximization of long-term value, with a better structure and a clearer value orientation.

[0157] The inventors also discovered that traditional recommendation systems built on user profiles, relying solely on surface-level data correlations for matching, are prone to the "information cocoon" effect. They fail to effectively distinguish between genuine causal relationships and statistical pseudo-correlation between user behavior variables, resulting in recommendation strategies that merely cater to current user preferences. They lack the ability to uncover deeper user needs and proactively guide them, and are also deficient in effective intervention methods and counterfactual reasoning capabilities, hindering the deep transformation of the cognitive value of data assets. To address these issues, this invention introduces causal discovery and counterfactual reasoning techniques. First, it automatically identifies the genuine causal structure between user behavior and long-term value from the full range of observational data. Based on this, a counterfactual reasoning model is constructed. By quantitatively evaluating the conditional average intervention effect of different intervention strategies on macro-value indicators, the optimal intervention strategy that can break the information cocoon and guide users towards a path that enhances macro-value indicators is accurately selected. Ultimately, based on causal intervention, a paradigm shift from passively predicting preferences to proactively shaping value is achieved, deeply mining and releasing the cognitive value inherent in data assets.

[0158] In some embodiments, Figure 9 The flowchart shown is a fifth embodiment of the user profile determination method provided by the present invention. Figure 10 The illustration shows a data flow diagram provided by the present invention, such as... Figure 9 and Figure 10 As shown, step 280 above may include the following sub-steps:

[0159] Step 510: Based on the causal structure corresponding to the target user and the user profile of the target user, perform matching processing in the preset strategy library to obtain candidate intervention strategies to be implemented.

[0160] For example, the causal structure is obtained by the causal discovery module using a causal discovery algorithm to perform causal identification processing on the observed data of the target user. It is a structured causal graph that can characterize the true causal relationship between intervention strategies, user profiles, and macro-value indicators, and is used to distinguish between true causal effects and spurious correlations. The causal algorithm can be, for example, the Peter-Clark algorithm, the Fast Causal Inference (FCI) algorithm, etc., and is not limited to this embodiment of the invention. The causal structure can be in the form of a Directed Acyclic Graph (DAG).

[0161] Observational data is a complete historical data set used to uncover causal relationships, including user profiles, historical behavioral data, log data of implemented intervention strategies, and feedback data of implemented intervention strategies.

[0162] The pre-defined strategy library contains a collection of intervention plans aimed at improving macro-level user value indicators. The matching process combines user profile features with effective causal paths in the causal structure to select intervention plans with potential causal improvement effects from the strategy library.

[0163] Candidate intervention strategies are a set of intervention strategies that have been initially screened and are likely to have a positive impact on the target users.

[0164] In one example, the user profile features of the target user are mapped to the corresponding causal structure to locate the interventionable nodes and paths that are causally related to macro-value indicators. Based on the causal path, feature matching and rule matching are performed in a preset strategy library to select intervention strategies that are suitable for the user profile and are located on an effective causal link, forming a set of candidate intervention strategies to be implemented.

[0165] Step 520: Based on the reasoning model of the target user, calculate the conditional average intervention effect of the candidate intervention strategies to be implemented.

[0166] For example, the inference model is used to infer the conditional average intervention effect after applying candidate intervention strategies.

[0167] The Conditional Average Treatment Effect (CATE) characterizes the degree to which a candidate intervention strategy improves macro-value indicators. In other words, given a user profile, the expected improvement in macro-value indicators relative to the state without intervention after applying a candidate intervention strategy is used to quantify the real value gain of the candidate intervention strategy.

[0168] In one example, the user profile of the target user, the candidate intervention strategies to be implemented, and the corresponding causal structure are input into the inference model. The inference model performs counterfactual inference to predict the macroeconomic value outcome after applying the candidate intervention strategy and the macroeconomic value outcome under the natural state. The difference between the two outcomes is taken as the conditional average intervention effect of the candidate intervention strategy.

[0169] Specifically, based on the counterfactual inference model in the inference model, the counterfactual value outcome corresponding to the candidate intervention strategy is predicted; based on the non-intervention inference model in the inference model, the factual value outcome corresponding to the candidate intervention strategy is determined; and the difference between the counterfactual value outcome and the factual value outcome is determined as the conditional average intervention effect of the candidate intervention strategy.

[0170] The inference model comprises two branches: a counterfactual inference model and a no-intervention inference model. The counterfactual inference model simulates the changes in the target user's future macro-value indicators after applying candidate intervention strategies. The no-intervention inference model simulates the future macro-value indicator results under the user's natural behavior without any intervention, i.e., the factual value results. The conditional average intervention effect is the difference between the counterfactual value results and the factual value results, used to quantify the actual improvement effect of the candidate intervention strategies on the macro-value indicators.

[0171] For example, the user profile of the target user, ecological association characteristics, and candidate intervention strategies are input into a counterfactual reasoning model. The model performs counterfactual inference based on the causal structure to predict the future macroeconomic value index after implementing the intervention strategy, i.e., the counterfactual value result. At the same time, the same user profile and ecological association characteristics are input into a no-intervention reasoning model. The model performs simulation inference under the constraint of maintaining natural behavioral evolution and without imposing any external intervention to obtain the future macroeconomic value index without intervention, i.e., the factual value result. The counterfactual value result is subtracted from the factual value result, and the calculated difference is used as the conditional average intervention effect corresponding to the candidate intervention strategy.

[0172] By using a dual-model parallel extrapolation approach, the future value results under the intervention state and the natural state are obtained separately. The real value enhancement effect of the strategy is quantified in the form of difference. This approach can effectively eliminate the interference of value changes caused by user behavior trends, accurately separate the causal gain generated by the intervention strategy itself, and make the calculation of the conditional average intervention effect more objective and reliable.

[0173] Step 530: Based on the conditional average intervention effect of the candidate intervention strategies to be implemented, determine the personalized intervention strategy for the target user.

[0174] For example, a personalized intervention strategy is the optimal intervention plan selected from candidate strategies that can produce the greatest positive improvement on the target user's macro value indicators.

[0175] In one example, the conditional average intervention effects of each candidate intervention strategy are numerically sorted, and the candidate intervention strategy with the largest and positive effect value is selected as the personalized intervention strategy for the target user. If none of the candidate intervention strategies produce a positive effect, no intervention or maintaining the original strategy is selected.

[0176] For example, the causal structure is "price sensitivity". In the "Purchase Decision" process, if the candidate intervention strategy to be implemented is "reduce the user's price sensitivity by 10%", counterfactual reasoning is conducted for the target user. This counterfactual reasoning model can answer "what will happen if a certain action is taken", and obtain the counterfactual value result. By comparing the factual value result (the value result in the state without intervention) with the counterfactual value result (the value result after intervention), the conditional average intervention effect corresponding to different intervention strategies (such as personalized pricing, content incentives, etc.) is quantitatively evaluated, thereby selecting the optimal intervention strategy.

[0177] In another example, the electronic device also includes an intervention strategy generator. First, based on the positive or negative effect, effect strength, and effect confidence interval of the conditional average intervention effect of each candidate intervention strategy, the optimization direction and control range of each candidate intervention strategy are determined. At the same time, user retention rate, platform ecosystem health, and long-term value conversion efficiency are taken as optimization objectives. Through algorithm model, the massive combination of candidate intervention strategies is iteratively deduced and weighted. On the basis of balancing the trade-offs between the objectives, personalized intervention strategies that can maximize comprehensive value and meet the actual feasibility of implementation are selected.

[0178] Optionally, the effectiveness of the personalized intervention strategy can be evaluated to obtain feedback data on the implemented intervention strategy; based on the feedback data on the implemented intervention strategy, the causal structure and inference model of the target user can be updated.

[0179] Among them, the evaluation of implementation effect is the process of tracking and statistically analyzing the user's subsequent behavior and changes in value indicators after the personalized intervention strategy has been implemented.

[0180] The feedback data for implemented intervention strategies is a set of quantitative results obtained from the evaluation. This feedback data includes at least one of the following: macro-value indicators, retention rate, and conversion rate. Among them, the retention rate represents the proportion of users who continue to use the service within the preset period after the intervention, and the conversion rate represents the proportion of users who complete target behaviors such as payment, content publishing, and task completion.

[0181] The causal structure update is based on the correction of the original causal relationship diagram by adding or deleting causal edges and adjusting the causal strength and direction, so that the causal structure is more in line with the actual law of action.

[0182] The inference model update uses newly added intervention-feedback data as incremental samples to incrementally train or fine-tune the counterfactual inference model and the non-intervention inference model, thereby improving the accuracy of model value prediction and effect estimation.

[0183] For example, after the personalized intervention strategy for the target user is implemented, the value assessment and feedback module continuously collects data on changes in macro-value indicators, retention status, and conversion behavior of the user within a preset observation period through A / B testing, long-term value tracking, and other methods. The above data is statistically and normally processed to form structured feedback data of the implemented intervention strategy. This feedback data is input into the causal discovery module as new observation data. Combined with historical observation data, causal identification and structural calibration are performed again to update the causal path and causal association strength in the causal structure. At the same time, the newly added intervention strategy, user profile, and corresponding feedback data are used as incremental training samples to perform incremental iterative optimization of the counterfactual reasoning model and the no-intervention reasoning model. The model parameters are adjusted to reduce the error between the predicted value and the actual feedback value, thus completing the synchronous update of the causal structure and the reasoning model.

[0184] Through the above steps, the paradigm of recommendation systems is transformed from passively "predicting user preferences" to proactively "taking causal interventions to maximize long-term value." By introducing causal discovery and counterfactual reasoning, the "information cocoon" problem commonly found in recommendation systems is effectively broken, upgrading the service paradigm from passive recommendation to proactive value creation. The optimal personalized intervention strategy that best guides users to migrate to high-value paths (improving macro-value indicators) is proactively selected, and a closed-loop feedback mechanism is formed by combining the feedback data after the intervention is implemented. This continuously optimizes the causal structure and counterfactual reasoning model, ultimately achieving a fundamental shift in the recommendation paradigm from simply predicting user preferences to proactively shaping long-term user value.

[0185] In this embodiment, candidate intervention strategies with causal feasibility are first screened based on the user's causal structure and user profile. Then, a dedicated inference model is used to calculate the conditional average intervention effect corresponding to each candidate intervention strategy. Finally, the optimal personalized intervention strategy is determined based on the magnitude of the conditional average intervention effect. This approach, based on real causal relationships rather than simple data correlation, can effectively eliminate interfering strategies that only have correlation but no actual improvement effect. This ensures that the selected intervention strategy has a stable and reliable value improvement effect, avoids the strategy misselection problem caused by traditional correlation-based matching, significantly improves the accuracy and effectiveness of personalized intervention, and ensures that the intervention operation can truly act on the causal path of value improvement.

[0186] Figure 11 A schematic diagram of a system architecture provided by the present invention is shown, such as... Figure 11 As shown, it is a four-layer collaborative cognitive architecture including a quantum perception layer, a neural symbol fusion layer, a cognitive image layer, and a causal service layer.

[0187] First, the quantum perception layer receives multimodal behavioral data input. The quantum state encoder maps the fused multimodal behavioral features into a normalized complex state vector in Hilbert space. The dynamic evolution of the user's intention state is completed through superposition and collapse calculations. Then, the intention state information, which represents the real-time probability distribution of the user's intention, is output through uncertainty quantization. The behavioral data and intention state information are then input together into the neural symbol fusion layer.

[0188] The neural symbol fusion layer achieves parallel processing of data-driven and knowledge reasoning through neural network channels and symbolic reasoning channels. The semantic alignment and conflict resolution center completes the semantic alignment and conflict resolution of the outputs of the two types of channels, realizing the deep fusion of multi-source cognitive results and outputting a micro-profile of the current time period.

[0189] The cognitive image layer takes the micro-portrait of the current time period output by the neural symbol fusion layer as input, and completes the cross-scale mapping from micro-feature vectors to macro-value indicators through the scale transformation function (macro-value prediction model). It generates a macro-portrait containing macro-value indicators such as user lifetime value, and updates the value contribution map to quantify the network effect value and ecological contribution value of users in the social network. The value contribution map data is then fed back to the causal service layer.

[0190] Based on the value contribution map and full-scale observation data from the cognitive image layer, the causal service layer automatically identifies the true causal graph structure between user behavior variables through a causal discovery algorithm. On this basis, a counterfactual reasoning model is constructed to generate multiple sets of candidate intervention strategies. The conditional average intervention effect of each candidate strategy is calculated, and the personalized intervention strategy that maximizes long-term user value is selected. Simultaneously, feedback data of implemented personalized intervention strategies is collected through value assessment and feedback. This feedback data is then output as feedback signals to the quantum perception layer, the neural symbol fusion layer, and the cognitive image layer, respectively, achieving dynamic calibration and iterative optimization across the entire chain. Ultimately, this realizes a paradigm shift from passive user intent perception to proactive causal intervention, and from micro-level instantaneous intent modeling to macro-level long-term value shaping, enabling cross-scale accurate measurement of user profiles and proactive value creation through personalized services.

[0191] In summary, existing user profiling and personalized recommendation technologies suffer from multiple core defects, failing to meet the needs of deep user understanding and long-term value creation. Traditional probabilistic models can only depict deterministic probability distributions, unable to effectively handle the ambiguity, superposition, and dynamic uncertainty of user intentions, and struggling to quantify the complex psychological states of users such as "indecision" during the selection process. Furthermore, recommendation systems built on historical data association rely solely on superficial correlations for matching, easily fostering "information cocoons," failing to distinguish between genuine causal relationships and statistically spurious correlations between user behavior variables, lacking proactive intervention and counterfactual reasoning capabilities, and hindering the upgrade from passive recommendation to proactive value creation. In addition, data-driven neural networks face deep integration barriers with symbolic business rules, domain common sense, and other prior knowledge, resulting in a black box and poor interpretability in the system's decision-making process, failing to provide users with transparent and traceable decision-making basis. Moreover, existing value assessments are mostly limited to short-term clicks and conversions, lacking effective measurement and positive incentive mechanisms for macro-values ​​such as long-term user lifetime value and ecosystem network contribution value, easily falling into short-term traffic traps and overdrawing long-term user value.

[0192] Therefore, this invention proposes a four-layer cognitive architecture that integrates quantum theory, neurosymbolic learning, and causal reasoning. Based on a quantum perception layer, it models user intent as a quantum superposition state in Hilbert space using a quantum probabilistic graphical model and performs collapse measurement to quantify the uncertainty of user intent, accurately characterizing the user's implicit intent and dynamic choice state. Simultaneously, a neurosymbolic fusion layer constructs a dual-channel collaborative architecture of neural network and symbolic reasoning channels. A semantic alignment and conflict resolution center enables deep integration of data-driven feature learning with symbolic business rules, domain common sense, and other prior knowledge, ensuring the interpretability of the reasoning process and obtaining a microscopic profile. Based on this, a cognitive profile layer is then constructed. The dual-track profiling system, which combines micro-behavioral particles and macro-value entities, achieves cross-scale mapping from micro-feature vectors of real-time intent states at the second level to long-term macro-value indicators at the quarterly / annual level through a learnable scale transformation function. This generates a dynamic panoramic cognitive profile that integrates instantaneous micro-behavior and long-term macro-value. Finally, at the causal service layer, based on causal discovery algorithms, it automatically identifies the true causal structure between user behavior variables from the full range of observation data. Combined with counterfactual reasoning models, it generates multiple sets of candidate intervention strategies to be implemented. By quantitatively evaluating the conditional average intervention effect of each strategy, it selects the optimal personalized intervention plan, realizing a technological paradigm shift from "association recommendation" to "causal intervention" and from "static profiling" to "cross-scale cognition".

[0193] This invention, through the deep synergy of the aforementioned four-layer cognitive architecture, achieves multi-dimensional performance breakthroughs and paradigm innovations compared to existing technologies. It can accurately model the uncertainty and implicit intentions of user intent that traditional technologies cannot depict. Furthermore, theoretical analysis and simulation experiments have verified that the accuracy of user intent recognition is improved by more than 60%. At the same time, the neural symbolic fusion architecture makes the recommendation decision-making process transparent and traceable, providing a complete reasoning path explanation based on business rules and domain common sense, effectively enhancing user trust and significantly improving decision interpretability. In addition, the causal intervention model can proactively guide users to migrate to high-value paths, which is expected to increase user lifetime value by more than 30%, maximizing the long-term value of users. At the same time, the value contribution graph quantifies the ecological value and constructs a positive incentive mechanism. Finally, the deep integration of quantum heuristic intent modeling and neural symbolic learning has significant foresight and creativity in its technical approach, building a technological paradigm barrier.

[0194] Figure 12 A schematic diagram of an embodiment of the user profile determination device provided by the present invention is shown. Figure 12 As shown, the user profile determination device 600 includes: an acquisition module 610, a processing module 620, a determination module 630, and an integration module 640.

[0195] Module 610 is used to acquire the target user's behavior data in the current time period;

[0196] The processing module 620 is used to perform intent modeling and prediction processing on the behavioral data of the current time period to obtain a micro profile of the target user in the current time period; wherein, the micro profile represents the weight distribution of the intent components of the target user in the current time period.

[0197] The determination module 630 is used to determine the macro value indicators of the target user based on the preset macro value prediction model and the micro profile of the current period; wherein, the macro value prediction model is configured to learn a non-linear mapping from intent components to macro value indicators.

[0198] The integration module 640 is used to obtain the user profile of the target user based on the micro profile and macro value indicators of the current time period.

[0199] In one alternative embodiment, the processing module 620 is used for:

[0200] Intent modeling is performed on the behavioral data of the current time period to obtain intent state information; wherein, the intent state information represents the weight distribution of at least one intent base state; the intent base state represents the user's potential interest points;

[0201] The intent state information and the behavior data of the current time period are input into a preset neural network model to obtain the intent prediction information output by the neural network model; wherein, the intent prediction information represents the predicted intent component and the initial weight distribution.

[0202] Based on intent state information, behavioral data for the current time period, and preset business rule information, intent assertion information is determined; wherein, intent assertion information represents the assertion result of predicting intent components;

[0203] Semantic alignment and conflict arbitration are performed on intent prediction information and intent assertion information to obtain a micro-profile of the target user in the current time period.

[0204] In one alternative embodiment, the processing module 620 is used for:

[0205] Determine the time evolution operator based on the behavioral data of the current time period;

[0206] The intent superposition state information of the target user in the previous time period is processed by the time evolution operator to obtain the intent superposition state information in the current time period; wherein, the intent superposition state information is a linear superposition representation of at least one intent base state.

[0207] Based on the observation operators corresponding to the behavioral data of the current time period, the intent superposition information of the current time period is collapsed to obtain the intent state information.

[0208] In an alternative embodiment, the device further includes a generation module for:

[0209] When preset conditions are met, a value gradient signal is generated based on the micro-profile of the target user in the previous time period; the value gradient signal is used to adjust the weight distribution of each intent component in the micro-profile in the current time period; the preset conditions include reaching a preset period or triggering a preset key node.

[0210] In one alternative approach, a module is generated for:

[0211] Based on the pre-defined intelligent agent, the micro-profile of the previous period is simulated and predicted to obtain the first future macro-value index corresponding to the micro-profile of the previous period and the second future macro-value index corresponding to the adjusted micro-profile; wherein, the adjusted micro-profile is obtained by adjusting the weight of the target intent component in the micro-profile of the previous period.

[0212] A value gradient signal is generated based on the first and second future macroeconomic value indicators.

[0213] In one alternative approach, a module is generated for:

[0214] Based on the pre-set graph convolutional network and the pre-set value contribution graph, the ecological association characteristics of the target user are determined; among them, the value contribution graph is a heterogeneous graph constructed based on the target user's historical behavior data, social relationship data, and content dissemination link data.

[0215] The micro-profile and ecological association features from the previous time period are input into the intelligent agent to simulate the first behavioral data of the target user in the future time period under the intention represented by the micro-profile in the previous time period; and the adjusted micro-profile and ecological association features are input into the intelligent agent to simulate the second behavioral data of the target user in the future time period under the intention represented by the adjusted micro-profile.

[0216] Based on the intelligent agent, the intention modeling and prediction processing of the first behavioral data in the future time period are performed to obtain the first micro profile of the future time period; and based on the intelligent agent, the intention modeling and prediction processing of the second behavioral data in the future time period are performed to obtain the second micro profile of the future time period.

[0217] Based on the intelligent agent, a first future macroeconomic value indicator is obtained according to the first microeconomic profile and macroeconomic value prediction model for the future period; and based on the intelligent agent, a second future macroeconomic value indicator is obtained according to the second microeconomic profile and macroeconomic value prediction model for the future period.

[0218] In one alternative embodiment, the processing module 620 is used for:

[0219] Semantic alignment is performed on the intent prediction information and intent assertion information to obtain aligned intent prediction information and aligned intent assertion information;

[0220] Based on the value gradient signal in the preset arbitration logic parameters, the aligned intent prediction information and the aligned intent assertion information are arbitrated to obtain the micro profile of the target user in the current time period.

[0221] In one alternative embodiment, the processing module 620 is used for:

[0222] Based on the value gradient signal, the system parameters are optimized to obtain the optimized system parameters; among them, the system parameters include at least one of the quantum parameters in intention modeling, the model parameters of the neural network model, and the arbitration logic parameters of conflict arbitration.

[0223] Based on the optimized system parameters, intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period.

[0224] In an alternative embodiment, the device further includes a matching module for:

[0225] Based on the user profile of the target user, a matching process is performed in a pre-set strategy library to determine the personalized intervention strategy for the target user; wherein, the personalized intervention strategy is an intervention strategy used to improve the macro value indicators of the target user.

[0226] Implement personalized intervention strategies.

[0227] In one alternative approach, the matching module is used for:

[0228] Based on the causal structure corresponding to the target user and the user profile of the target user, a matching process is performed in a preset strategy library to obtain candidate intervention strategies to be implemented. The causal structure is obtained by using a causal discovery algorithm to perform causal identification processing on the observation data of the target user. The observation data includes user profile, historical behavior data, log data of implemented intervention strategies, and feedback data of implemented intervention strategies.

[0229] Based on the inference model of the target user, the conditional average intervention effect of the candidate intervention strategy is calculated; whereby the conditional average intervention effect characterizes the degree of improvement of the macro value indicator by the candidate intervention strategy; the inference model is used to infer the conditional average intervention effect after the candidate intervention strategy is applied.

[0230] Based on the conditional average intervention effect of candidate intervention strategies, personalized intervention strategies for target users are determined.

[0231] In one alternative approach, the matching module is used for:

[0232] Based on the counterfactual reasoning model in the reasoning model, predict the counterfactual value outcome corresponding to the candidate intervention strategies to be implemented;

[0233] Based on the non-intervention reasoning model in the reasoning model, the factual value outcome corresponding to the candidate intervention strategies to be implemented is determined.

[0234] The difference between the counterfactual value outcome and the factual value outcome is determined as the conditional average intervention effect of the candidate intervention strategy.

[0235] In one alternative approach, the matching module is used for:

[0236] The effectiveness of the personalized intervention strategy is evaluated to obtain feedback data on the implemented intervention strategy; the feedback data on the implemented intervention strategy includes at least one of the following: macro value indicators, retention rate, and conversion rate.

[0237] Based on feedback data from implemented intervention strategies, update the causal structure and inference model for target users.

[0238] As can be seen from the above, the user profile determination device provided in this embodiment of the invention can accurately quantify and depict the vague implicit intentions of users from the bottom layer. At the same time, it introduces a measurement dimension of long-term macro value, which breaks through the one-sidedness of traditional user profiles that only focus on explicit behavior and short-term value. It greatly improves the comprehensiveness and accuracy of user profiles, allowing the profiles to truly reflect the deep needs and long-term value attributes of users, and provides more reliable data support and decision-making basis for subsequent precise services, value guidance and ecosystem operation based on user cognition.

[0239] Figure 13 The diagram shows a structural schematic of an embodiment of the electronic device provided by the present invention. The specific embodiments of the present invention do not limit the specific implementation of the electronic device.

[0240] like Figure 13 As shown, the electronic device may include: a processor 702, a communications interface 704, a memory 706, and a communications bus 708.

[0241] The processor 702, communication interface 704, and memory 706 communicate with each other via communication bus 708. Communication interface 704 is used to communicate with other network elements such as clients or other servers. The processor 702 executes program 710, specifically performing the relevant steps in the above-described embodiment of the user profile determination method.

[0242] Specifically, program 710 may include program code, which includes computer-executable instructions.

[0243] Processor 702 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The electronic device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0244] Memory 706 is used to store program 710. Memory 706 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0245] Specifically, program 710 can be called by processor 702 to cause the electronic device to perform the following operations:

[0246] Obtain the target user's behavioral data for the current time period;

[0247] Intent modeling and prediction processing are performed on behavioral data for the current time period to obtain a micro-profile of the target user for the current time period; where the micro-profile represents the weight distribution of the intent components of the target user for the current time period;

[0248] Based on a pre-defined macro-value prediction model and a micro-profile of the current period, macro-value indicators for target users are determined; the macro-value prediction model is configured to learn a non-linear mapping from intent components to macro-value indicators.

[0249] Based on the micro-profiles and macro-value indicators of the current time period, a user profile of the target user is obtained.

[0250] In one alternative approach, intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period, including:

[0251] Intent modeling is performed on the behavioral data of the current time period to obtain intent state information; wherein, the intent state information represents the weight distribution of at least one intent base state; the intent base state represents the user's potential interest points;

[0252] The intent state information and the behavior data of the current time period are input into a preset neural network model to obtain the intent prediction information output by the neural network model; wherein, the intent prediction information represents the predicted intent component and the initial weight distribution.

[0253] Based on intent state information, behavioral data for the current time period, and preset business rule information, intent assertion information is determined; wherein, intent assertion information represents the assertion result of predicting intent components;

[0254] Semantic alignment and conflict arbitration are performed on intent prediction information and intent assertion information to obtain a micro-profile of the target user in the current time period.

[0255] In one alternative approach, intent modeling is performed on the behavioral data of the current time period to obtain intent state information, including:

[0256] Determine the time evolution operator based on the behavioral data of the current time period;

[0257] The intent superposition state information of the target user in the previous time period is processed by the time evolution operator to obtain the intent superposition state information in the current time period; wherein, the intent superposition state information is a linear superposition representation of at least one intent base state.

[0258] Based on the observation operators corresponding to the behavioral data of the current time period, the intent superposition information of the current time period is collapsed to obtain the intent state information.

[0259] In one alternative approach, the method further includes:

[0260] When preset conditions are met, a value gradient signal is generated based on the micro-profile of the target user in the previous time period; the value gradient signal is used to adjust the weight distribution of each intent component in the micro-profile in the current time period; the preset conditions include reaching a preset period or triggering a preset key node.

[0261] In one alternative approach, a value gradient signal is generated based on the target user's micro-profile from the previous time period, including:

[0262] Based on the pre-defined intelligent agent, the micro-profile of the previous period is simulated and predicted to obtain the first future macro-value index corresponding to the micro-profile of the previous period and the second future macro-value index corresponding to the adjusted micro-profile; wherein, the adjusted micro-profile is obtained by adjusting the weight of the target intent component in the micro-profile of the previous period.

[0263] A value gradient signal is generated based on the first and second future macroeconomic value indicators.

[0264] In one optional approach, based on a pre-defined intelligent agent, a simulation prediction process is performed on the micro-profile from the previous time period to obtain a first future macro-value indicator corresponding to the micro-profile from the previous time period and a second future macro-value indicator corresponding to the adjusted micro-profile, including:

[0265] Based on the pre-set graph convolutional network and the pre-set value contribution graph, the ecological association characteristics of the target user are determined; among them, the value contribution graph is a heterogeneous graph constructed based on the target user's historical behavior data, social relationship data, and content dissemination link data.

[0266] The micro-profile and ecological association features from the previous time period are input into the intelligent agent to simulate the first behavioral data of the target user in the future time period under the intention represented by the micro-profile in the previous time period; and the adjusted micro-profile and ecological association features are input into the intelligent agent to simulate the second behavioral data of the target user in the future time period under the intention represented by the adjusted micro-profile.

[0267] Based on the intelligent agent, the intention modeling and prediction processing of the first behavioral data in the future time period are performed to obtain the first micro profile of the future time period; and based on the intelligent agent, the intention modeling and prediction processing of the second behavioral data in the future time period are performed to obtain the second micro profile of the future time period.

[0268] Based on the intelligent agent, a first future macroeconomic value indicator is obtained according to the first microeconomic profile and macroeconomic value prediction model for the future period; and based on the intelligent agent, a second future macroeconomic value indicator is obtained according to the second microeconomic profile and macroeconomic value prediction model for the future period.

[0269] In one alternative approach, semantic alignment and conflict arbitration are performed on the intent prediction information and intent assertion information to obtain a micro-profile of the target user in the current time period, including:

[0270] Semantic alignment is performed on the intent prediction information and intent assertion information to obtain aligned intent prediction information and aligned intent assertion information;

[0271] Based on the value gradient signal in the preset arbitration logic parameters, the aligned intent prediction information and the aligned intent assertion information are arbitrated to obtain the micro profile of the target user in the current time period.

[0272] In one alternative approach, the method further includes:

[0273] Based on the value gradient signal, the system parameters are optimized to obtain the optimized system parameters; among them, the system parameters include at least one of the quantum parameters in intention modeling, the model parameters of the neural network model, and the arbitration logic parameters of conflict arbitration.

[0274] Intent modeling and prediction are performed on behavioral data for the current time period to obtain a micro-profile of the target user for the current time period, including:

[0275] Based on the optimized system parameters, intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period.

[0276] In one alternative approach, the method further includes:

[0277] Based on the user profile of the target user, a matching process is performed in a pre-set strategy library to determine the personalized intervention strategy for the target user; wherein, the personalized intervention strategy is an intervention strategy used to improve the macro value indicators of the target user.

[0278] Implement personalized intervention strategies.

[0279] In one alternative approach, based on the target user's user profile, a matching process is performed in a pre-defined strategy library to determine the target user's personalized intervention strategy, including:

[0280] Based on the causal structure corresponding to the target user and the user profile of the target user, a matching process is performed in a preset strategy library to obtain candidate intervention strategies to be implemented. The causal structure is obtained by using a causal discovery algorithm to perform causal identification processing on the observation data of the target user. The observation data includes user profile, historical behavior data, log data of implemented intervention strategies, and feedback data of implemented intervention strategies.

[0281] Based on the inference model of the target user, the conditional average intervention effect of the candidate intervention strategy is calculated; whereby the conditional average intervention effect characterizes the degree of improvement of the macro value indicator by the candidate intervention strategy; the inference model is used to infer the conditional average intervention effect after the candidate intervention strategy is applied.

[0282] Based on the conditional average intervention effect of candidate intervention strategies, personalized intervention strategies for target users are determined.

[0283] In one alternative approach, based on the inference model of the target user, the conditional average intervention effect of the candidate intervention strategies to be implemented is calculated, including:

[0284] Based on the counterfactual reasoning model in the reasoning model, predict the counterfactual value outcome corresponding to the candidate intervention strategies to be implemented;

[0285] Based on the non-intervention reasoning model in the reasoning model, the factual value outcome corresponding to the candidate intervention strategies to be implemented is determined.

[0286] The difference between the counterfactual value outcome and the factual value outcome is determined as the conditional average intervention effect of the candidate intervention strategy.

[0287] In one alternative approach, the method further includes:

[0288] The effectiveness of the personalized intervention strategy is evaluated to obtain feedback data on the implemented intervention strategy; the feedback data on the implemented intervention strategy includes at least one of the following: macro value indicators, retention rate, and conversion rate.

[0289] Based on feedback data from implemented intervention strategies, update the causal structure and inference model for target users.

[0290] As can be seen from the above, the electronic device provided by the embodiments of the present invention can accurately quantify and characterize the vague implicit intentions of users from the bottom layer. At the same time, it introduces a measurement dimension of long-term macro value, which breaks through the one-sidedness of traditional user profiles that only focus on explicit behavior and short-term value. It greatly improves the comprehensiveness and accuracy of user profiles, allowing the profiles to truly reflect the deep needs and long-term value attributes of users, and provides more reliable data support and decision-making basis for subsequent precise services, value guidance and ecosystem operation based on user cognition.

[0291] This invention provides a computer-readable storage medium storing at least one executable instruction. When the executable instruction is executed on an electronic device / user profile determination device, the electronic device / user profile determination device performs the user profile determination method in any of the above method embodiments.

[0292] Specifically, the executable instructions can be used to cause the electronic device / user profile determination device to perform the following operations:

[0293] Obtain the target user's behavioral data for the current time period;

[0294] Intent modeling and prediction processing are performed on behavioral data for the current time period to obtain a micro-profile of the target user for the current time period; where the micro-profile represents the weight distribution of the intent components of the target user for the current time period;

[0295] Based on a pre-defined macro-value prediction model and a micro-profile of the current period, macro-value indicators for target users are determined; the macro-value prediction model is configured to learn a non-linear mapping from intent components to macro-value indicators.

[0296] Based on the micro-profiles and macro-value indicators of the current time period, a user profile of the target user is obtained.

[0297] In one alternative approach, intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period, including:

[0298] Intent modeling and observation processing are performed on behavioral data in the current time period to obtain intent state information; wherein, intent state information represents the weight distribution of at least one intent base state; the intent base state represents the user's potential interest points;

[0299] The intent state information and the behavior data of the current time period are input into a preset neural network model to obtain the intent prediction information output by the neural network model; wherein, the intent prediction information represents the predicted intent component and the initial weight distribution.

[0300] Based on intent state information, behavioral data for the current time period, and preset business rule information, intent assertion information is determined; wherein, intent assertion information represents the assertion result of predicting intent components;

[0301] Semantic alignment and conflict arbitration are performed on intent prediction information and intent assertion information to obtain a micro-profile of the target user in the current time period.

[0302] In one alternative approach, intent modeling is performed on the behavioral data of the current time period to obtain intent state information, including:

[0303] Determine the time evolution operator based on the behavioral data of the current time period;

[0304] The intent superposition state information of the target user in the previous time period is processed by the time evolution operator to obtain the intent superposition state information in the current time period; wherein, the intent superposition state information is a linear superposition representation of at least one intent base state.

[0305] Based on the observation operators corresponding to the behavioral data of the current time period, the intent superposition information of the current time period is collapsed to obtain the intent state information.

[0306] In one alternative approach, the method further includes:

[0307] When preset conditions are met, a value gradient signal is generated based on the micro-profile of the target user in the previous time period; the value gradient signal is used to adjust the weight distribution of each intent component in the micro-profile in the current time period; the preset conditions include reaching a preset period or triggering a preset key node.

[0308] In one alternative approach, a value gradient signal is generated based on the target user's micro-profile from the previous time period, including:

[0309] Based on the pre-defined intelligent agent, the micro-profile of the previous period is simulated and predicted to obtain the first future macro-value index corresponding to the micro-profile of the previous period and the second future macro-value index corresponding to the adjusted micro-profile; wherein, the adjusted micro-profile is obtained by adjusting the weight of the target intent component in the micro-profile of the previous period.

[0310] A value gradient signal is generated based on the first and second future macroeconomic value indicators.

[0311] In one optional approach, based on a pre-defined intelligent agent, a simulation prediction process is performed on the micro-profile from the previous time period to obtain a first future macro-value indicator corresponding to the micro-profile from the previous time period and a second future macro-value indicator corresponding to the adjusted micro-profile, including:

[0312] Based on the pre-set graph convolutional network and the pre-set value contribution graph, the ecological association characteristics of the target user are determined; among them, the value contribution graph is a heterogeneous graph constructed based on the target user's historical behavior data, social relationship data, and content dissemination link data.

[0313] The micro-profile and ecological association features from the previous time period are input into the intelligent agent to simulate the first behavioral data of the target user in the future time period under the intention represented by the micro-profile in the previous time period; and the adjusted micro-profile and ecological association features are input into the intelligent agent to simulate the second behavioral data of the target user in the future time period under the intention represented by the adjusted micro-profile.

[0314] Based on the intelligent agent, the intention modeling and prediction processing of the first behavioral data in the future time period are performed to obtain the first micro profile of the future time period; and based on the intelligent agent, the intention modeling and prediction processing of the second behavioral data in the future time period are performed to obtain the second micro profile of the future time period.

[0315] Based on the intelligent agent, a first future macroeconomic value indicator is obtained according to the first microeconomic profile and macroeconomic value prediction model for the future period; and based on the intelligent agent, a second future macroeconomic value indicator is obtained according to the second microeconomic profile and macroeconomic value prediction model for the future period.

[0316] In one alternative approach, semantic alignment and conflict arbitration are performed on the intent prediction information and intent assertion information to obtain a micro-profile of the target user in the current time period, including:

[0317] Semantic alignment is performed on the intent prediction information and intent assertion information to obtain aligned intent prediction information and aligned intent assertion information;

[0318] Based on the value gradient signal in the preset arbitration logic parameters, the aligned intent prediction information and the aligned intent assertion information are arbitrated to obtain the micro profile of the target user in the current time period.

[0319] In one alternative approach, the method further includes:

[0320] Based on the value gradient signal, the system parameters are optimized to obtain the optimized system parameters; among them, the system parameters include at least one of the quantum parameters in intention modeling, the model parameters of the neural network model, and the arbitration logic parameters of conflict arbitration.

[0321] Intent modeling and prediction are performed on behavioral data for the current time period to obtain a micro-profile of the target user for the current time period, including:

[0322] Based on the optimized system parameters, intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period.

[0323] In one alternative approach, the method further includes:

[0324] Based on the user profile of the target user, a matching process is performed in a pre-set strategy library to determine the personalized intervention strategy for the target user; wherein, the personalized intervention strategy is an intervention strategy used to improve the macro value indicators of the target user.

[0325] Implement personalized intervention strategies.

[0326] In one alternative approach, based on the target user's user profile, a matching process is performed in a pre-defined strategy library to determine the target user's personalized intervention strategy, including:

[0327] Based on the causal structure corresponding to the target user and the user profile of the target user, a matching process is performed in a preset strategy library to obtain candidate intervention strategies to be implemented. The causal structure is obtained by using a causal discovery algorithm to perform causal identification processing on the observation data of the target user. The observation data includes user profile, historical behavior data, log data of implemented intervention strategies, and feedback data of implemented intervention strategies.

[0328] Based on the inference model of the target user, the conditional average intervention effect of the candidate intervention strategy is calculated; whereby the conditional average intervention effect characterizes the degree of improvement of the macro value indicator by the candidate intervention strategy; the inference model is used to infer the conditional average intervention effect after the candidate intervention strategy is applied.

[0329] Based on the conditional average intervention effect of candidate intervention strategies, personalized intervention strategies for target users are determined.

[0330] In one alternative approach, based on the inference model of the target user, the conditional average intervention effect of the candidate intervention strategies to be implemented is calculated, including:

[0331] Based on the counterfactual reasoning model in the reasoning model, predict the counterfactual value outcome corresponding to the candidate intervention strategies to be implemented;

[0332] Based on the non-intervention reasoning model in the reasoning model, the factual value outcome corresponding to the candidate intervention strategies to be implemented is determined.

[0333] The difference between the counterfactual value outcome and the factual value outcome is determined as the conditional average intervention effect of the candidate intervention strategy.

[0334] In one alternative approach, the method further includes:

[0335] The effectiveness of the personalized intervention strategy is evaluated to obtain feedback data on the implemented intervention strategy; the feedback data on the implemented intervention strategy includes at least one of the following: macro value indicators, retention rate, and conversion rate.

[0336] Based on feedback data from implemented intervention strategies, update the causal structure and inference model for target users.

[0337] As can be seen from the above, the computer-readable storage medium provided in the embodiments of the present invention stores at least one executable instruction. When the executable instruction runs on the electronic device / user profile determination device, it can accurately quantify the vague implicit intentions of users from the bottom layer. At the same time, it introduces a measurement dimension of long-term macro value, which breaks through the one-sidedness of traditional user profiles that only focus on explicit behavior and short-term value. It greatly improves the comprehensiveness and accuracy of user profiles, allowing the profiles to truly reflect the deep needs and long-term value attributes of users, and provides more reliable data support and decision-making basis for subsequent precise services, value guidance and ecosystem operation based on user cognition.

[0338] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Furthermore, the embodiments of this invention are not directed to any particular programming language.

[0339] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. Similarly, for the sake of brevity and to aid in understanding one or more aspects of the invention, in the description of exemplary embodiments of the invention above, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of the invention.

[0340] Those skilled in the art will understand that the modules in the device of the embodiment can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiment can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

[0341] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A method for determining a user profile, characterized in that, include: Obtain the target user's behavioral data for the current time period; Intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period; wherein, the micro-profile represents the weight distribution of the intent components of the target user in the current time period; Based on a preset macro-value prediction model and the micro-profile of the current time period, the macro-value indicators of the target user are determined; wherein, the macro-value prediction model is configured to learn a non-linear mapping from intent components to macro-value indicators. Based on the micro-profile of the current time period and the macro-value indicators, the user profile of the target user is obtained.

2. The method according to claim 1, characterized in that, The process of performing intent modeling and prediction on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period includes: The behavioral data of the current time period is processed by intent modeling to obtain intent state information; wherein, the intent state information represents the weight distribution of at least one intent base state; the intent base state represents the user's potential interest points; The intent state information and the behavior data of the current time period are input into a preset neural network model to obtain the intent prediction information output by the neural network model; wherein, the intent prediction information represents the predicted intent component and the initial weight distribution; Based on the intent state information, the behavioral data of the current time period, and the preset business rule information, intent assertion information is determined; wherein, the intent assertion information represents the assertion result of predicting the intent component; Semantic alignment and conflict arbitration are performed on the intent prediction information and the intent assertion information to obtain a micro-profile of the target user in the current time period.

3. The method according to claim 2, characterized in that, The intention modeling process performed on the behavioral data of the current time period to obtain intention state information includes: Based on the behavioral data of the current time period, determine the time evolution operator; Based on the time evolution operator, the intent superposition state information of the target user in the previous time period is processed by time-series evolution to obtain the intent superposition state information of the current time period; wherein, the intent superposition state information is a linear superposition representation of at least one intent base state; Based on the observation operator corresponding to the behavior data of the current time period, the intent superposition state information of the current time period is collapsed to obtain the intent state information.

4. The method according to claim 2, characterized in that, The method further includes: When preset conditions are met, a value gradient signal is generated based on the micro-profile of the target user in the previous time period; wherein, the value gradient signal is used to adjust the weight distribution of each intent component in the micro-profile in the current time period; the preset conditions include reaching a preset period or triggering a preset key node.

5. The method according to claim 4, characterized in that, The step of generating a value gradient signal based on the target user's micro-profile in the previous time period includes: Based on a pre-defined intelligent agent, the micro-portrait of the previous time period is simulated and predicted to obtain a first future macro-value index corresponding to the micro-portrait of the previous time period and a second future macro-value index corresponding to the adjusted micro-portrait; wherein, the adjusted micro-portrait is obtained by adjusting the weight of the target intent component in the micro-portrait of the previous time period. The value gradient signal is generated based on the first future macroeconomic value indicator and the second future macroeconomic value indicator.

6. The method according to claim 5, characterized in that, The aforementioned intelligent agent performs simulation and prediction processing on the micro-profile of the previous time period to obtain a first future macro-value index corresponding to the micro-profile of the previous time period and a second future macro-value index corresponding to the adjusted micro-profile, including: Based on a preset graph convolutional network and a preset value contribution graph, the ecological association characteristics of the target user are determined; wherein, the value contribution graph is a heterogeneous graph constructed based on the target user's historical behavior data, social relationship data, and content dissemination link data; The micro-profile from the previous time period and the ecological association features are input into the intelligent agent to simulate the first behavioral data of the target user in the future time period under the intention represented by the micro-profile from the previous time period; and the adjusted micro-profile and the ecological association features are input into the intelligent agent to simulate the second behavioral data of the target user in the future time period under the intention represented by the adjusted micro-profile. Based on the intelligent agent, intent modeling and prediction processing are performed on the first behavioral data of the future time period to obtain the first micro-profile of the future time period; and based on the intelligent agent, intent modeling and prediction processing are performed on the second behavioral data of the future time period to obtain the second micro-profile of the future time period. Based on the intelligent agent, a first future macroeconomic value index is obtained according to the first microeconomic profile of the future period and the macroeconomic value prediction model; and based on the intelligent agent, a second future macroeconomic value index is obtained according to the second microeconomic profile of the future period and the macroeconomic value prediction model.

7. The method according to claim 4, characterized in that, The step of performing semantic alignment and conflict arbitration on the intent prediction information and the intent assertion information to obtain the micro-profile of the target user in the current time period includes: The intent prediction information and the intent assertion information are semantically aligned to obtain aligned intent prediction information and aligned intent assertion information. Based on the value gradient signal in the preset arbitration logic parameters, the aligned intent prediction information and the aligned intent assertion information are arbitrated to obtain the micro profile of the target user in the current time period.

8. The method according to claim 4, characterized in that, The method further includes: Based on the value gradient signal, the system parameters are optimized to obtain optimized system parameters; wherein, the system parameters include at least one of the quantum parameters in intention modeling, the model parameters of the neural network model, and the arbitration logic parameters of conflict arbitration. The process of performing intent modeling and prediction on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period includes: Based on the optimized system parameters, intent modeling and prediction processing are performed on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period.

9. The method according to any one of claims 1-8, characterized in that, The method further includes: Based on the user profile of the target user, a matching process is performed in a preset strategy library to determine the personalized intervention strategy for the target user; wherein, the personalized intervention strategy is an intervention strategy used to improve the macro value indicators of the target user. Implement the personalized intervention strategy.

10. The method according to claim 9, characterized in that, The step of determining a personalized intervention strategy for the target user by performing a matching process in a preset strategy library based on the target user's user profile includes: Based on the causal structure corresponding to the target user and the user profile of the target user, a matching process is performed in a preset strategy library to obtain candidate intervention strategies to be implemented; wherein, the causal structure is obtained by performing causal identification processing on the observation data of the target user using a causal discovery algorithm, and the observation data includes user profile, historical behavior data, log data of implemented intervention strategies, and feedback data of implemented intervention strategies; Based on the inference model of the target user, the conditional average intervention effect of the candidate intervention strategy is calculated; wherein, the conditional average intervention effect characterizes the degree of improvement of the macro value indicator by the candidate intervention strategy; the inference model is used to infer the conditional average intervention effect after the candidate intervention strategy is applied. Based on the conditional average intervention effect of the candidate intervention strategies to be implemented, a personalized intervention strategy for the target user is determined.

11. The method according to claim 10, characterized in that, The reasoning model based on the target user calculates the conditional average intervention effect of the candidate intervention strategies, including: Based on the counterfactual reasoning model in the inference model, predict the counterfactual value outcome corresponding to the candidate intervention strategy to be implemented; Based on the non-intervention reasoning model in the reasoning model, the factual value results corresponding to the candidate intervention strategies to be implemented are determined; The difference between the counterfactual value result and the factual value result is determined as the conditional average intervention effect of the candidate intervention strategy to be implemented.

12. The method according to claim 10, characterized in that, The method further includes: The effectiveness of the personalized intervention strategy is evaluated to obtain feedback data on the implemented intervention strategy; the feedback data on the implemented intervention strategy includes at least one of macro-value indicators, retention rate, and conversion rate. Based on the feedback data from the implemented intervention strategies, the causal structure and inference model of the target user are updated.

13. A device for determining a user profile, characterized in that, The device includes: The acquisition module is used to acquire the target user's behavioral data in the current time period; The processing module is used to perform intent modeling and prediction processing on the behavioral data of the current time period to obtain a micro-profile of the target user in the current time period; wherein, the micro-profile represents the weight distribution of the intent components of the target user in the current time period; The determination module is used to determine the macro value index of the target user based on a preset macro value prediction model and the micro profile of the current time period; wherein the macro value prediction model is configured to learn a nonlinear mapping from intent components to macro value indexes. The integration module is used to obtain the user profile of the target user based on the micro profile of the current time period and the macro value indicators.

14. An electronic device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation of the user profile determination method as described in any one of claims 1-12.

Images