A user preference analysis processing method, device and equipment

Abstract

The application discloses a user preference analysis processing method, device and equipment. Including: acquiring user behavior change data; determining a first intelligent agent configured to tend to affirm that the user preference has changed; performing semantic analysis on the user behavior change data by the first intelligent agent to generate user preference change argument data of this round; determining a second intelligent agent configured to tend to deny that the user preference has changed; performing semantic analysis on the user preference change argument data by the second intelligent agent to generate user preference change counterargument data and / or user preference change consensus reaching data, wherein the user preference change counterargument data is used for the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and the second intelligent agent to continue to counterargue or reach consensus in the next round; and determining the change of the user preference according to the processing of the first intelligent agent and the second intelligent agent for one or more rounds.

Description

Technical Field

[0001] This specification relates to the field of machine learning technology, and in particular to a user preference analysis and processing method, apparatus, and device. Background Technology

[0002] With the development of internet technology, many businesses can now be conveniently conducted online, and service providers are offering increasingly personalized services to improve user experience. To better provide these personalized services, service providers need to understand the preferences of different users, and then cater to those preferences to offer more precise and personalized services.

[0003] In practical applications, user preferences are not necessarily static, but may change with time, context, or user experience. Therefore, service providers also need to dynamically track changes in user preferences of different users.

[0004] In existing technologies, overly simplistic and crude strategies are often used to determine user preferences. A typical example is directly taking the user's current action as the user preference. While this approach has some degree of accuracy, it is also prone to misjudgment.

[0005] Therefore, there is a need for solutions that can help track changes in user preferences more reliably and accurately. Summary of the Invention

[0006] This specification provides one or more embodiments of a user preference analysis and processing method, apparatus, or device to solve the following technical problem: the need for a solution that helps to track changes in user preferences more reliably and accurately.

[0007] To solve the above-mentioned technical problems, one or more embodiments of this specification are implemented as follows: This specification provides a user preference analysis and processing method according to one or more embodiments, including: Obtain data on changes in user behavior; The first agent is determined to be configured to be inclined to believe that user preferences have changed. The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round. The second agent is configured to tend to deny changes in user preferences. The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data in the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. Based on one or more rounds of processing by the first agent and the second agent, the changes in user preferences are determined.

[0008] This specification provides a user preference analysis and processing apparatus according to one or more embodiments, comprising: The behavior change data acquisition module acquires user behavior change data. The first agent determination module determines the first agent that is configured to be inclined to affirm that the user's preferences have changed. The preference change argument generation module performs semantic analysis on the user behavior change data by the first intelligent agent to generate user preference change argument data for this round. The second agent determination module determines the second agent that is configured to tend to deny changes in user preferences. The preference change argument adversarial module performs semantic analysis on the user preference change argument data through the second intelligent agent to generate user preference change defense data and / or user preference change consensus data for the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. The user preference change determination module determines the changes in user preferences based on one or more rounds of processing by the first agent and the second agent.

[0009] This specification provides one or more embodiments of a user preference analysis and processing device, comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform: Obtain data on changes in user behavior; The first agent is determined to be configured to be inclined to believe that user preferences have changed. The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round. The second agent is configured to tend to deny changes in user preferences. The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data in the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. Based on one or more rounds of processing by the first agent and the second agent, the changes in user preferences are determined.

[0010] The above-mentioned at least one technical solution adopted in one or more embodiments of this specification can achieve the following beneficial effects: Through the adversarial dual-agent active debate architecture provided in this application, a first agent configured to be inclined to affirm that user preferences have changed acts as a challenger. Based on the current user behavior change data, as well as the semantic understanding and content generation capabilities of the large model, it initiates arguments related to the change in user preferences to challenge the first agent. A second agent configured to be inclined to deny that user preferences have changed acts as a defender to respond to the challenge of the first agent. Based on the semantic understanding and content generation capabilities of the large model, at least partial rebuttal and / or at least partial consensus are reached. In this way, one or more rounds of such dual-agent adversarial debate process can be carried out, which helps to approach the true state of user preferences in the tension of argumentation and questioning, better understand the current user preferences of the corresponding user, and can also use the newly generated user behavior change data to periodically or irregularly perform the same processing to achieve dynamic adversarial analysis over a longer period of time. Therefore, it helps to track changes in user preferences more reliably and accurately. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a user preference analysis and processing method provided in one or more embodiments of this specification; Figure 2 A flowchart illustrating a user preference change argument data processing scheme provided for one or more embodiments of this specification; Figure 3 Provided for one or more embodiments of this specification Figure 1A system architecture diagram of a specific implementation scheme of the method; Figure 4 A schematic diagram of a user preference analysis and processing device provided for one or more embodiments of this specification; Figure 5 This is a schematic diagram of the structure of a user preference analysis and processing device provided for one or more embodiments of this specification. Detailed Implementation

[0013] This specification provides a user preference analysis and processing method, apparatus, device, and storage medium through its embodiments.

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0015] In this application, changes in user preferences can be primarily understood as preference drift, specifically referring to the phenomenon that a user's interests, habits, or needs, which reflect preferences, change over time.

[0016] To track user preferences, a simplistic and crude strategy, as described in the background, might be implemented as a simple appending scheme. This involves directly using the user's current action as the user preference and appending new preferences directly to memory storage. If a conflict arises, the new preference simply overwrites the old one. The applicant has identified the following problems with this approach: lack of drift detection, failure to distinguish between occasional behavior and genuine preference changes, treating all new information equally; overwriting too quickly, with old preferences simply replaced by new information, unable to handle the coexistence of old and new preferences (e.g., different preferences in work and life scenarios); and lack of evidence evaluation, failing to assess the strength of evidence supporting a preference, treating preferences from a single action and those from hundreds of actions as equal. These problems could significantly increase the probability of misjudgment.

[0017] Based on this, the applicant may have considered further optimization and enhancement schemes, such as a scheme based on time decay and statistical thresholds. Specifically, this involves weighting user behavior by time decay while setting statistical thresholds (e.g., only considering a preference a occurrence after N occurrences) to comprehensively determine whether preferences have changed. However, the applicant also recognizes that this scheme has its problems: semantic blind spots, failing to understand the contextual meaning of behavior and struggling to identify scene transitions. For example, a user searching for "coffee shop A" three times is statistically a signal, but if all three searches occur in a business dinner scenario, semantically it is a scene-related behavior but not necessarily a change in user taste preferences; statistical methods cannot distinguish this. Threshold rigidity: different types of preferences change at different rates (e.g., dietary preferences may change quickly, while cultural inclinations change slowly), and fixed thresholds cannot adapt, requiring extensive manual tuning. Difficulty in distinguishing between occasional behavior and trend changes, and difficulty in identifying abrupt changes. Abrupt drift (e.g., moving) is not gradual accumulation and may generate a large number of new behaviors in a short period; the time decay model may react slowly due to the still high historical weights.

[0018] Due to the rise of large models (especially large language models) and their excellent analytical reasoning capabilities, applicants are also considering using large models to analyze and track user preferences.

[0019] One of the solutions the applicant attempted, termed the full recalculation solution, involved inputting user behavior data into a large model each period (e.g., daily or weekly), allowing the model to infer user preferences from scratch. The applicant further observed the following problems with this solution: unstable results; due to the randomness of the large model, the same behavioral data could produce significantly different preference descriptions in two inferences, leading users to perceive it as a split personality within the system; lack of historical anchors; the model struggles to distinguish between consistently stable preferences and newly emerging signals because it cannot see historical preference states; and sensitivity to noise; occasional behavior might be captured and amplified by the current inference, only to disappear in the next, causing preference flickering.

[0020] Similarly, referencing the full recalculation approach, a fixed window approach can be adopted, such as only considering the behavior of the most recent N days. In this case, the problem is that if the window is too short, there will be too much noise leading to overfitting, while if it is too long, the response will be slow leading to underfitting, and it will be difficult to adapt to different drift rates.

[0021] Another approach attempted by the applicant, termed a concept drift detection scheme, aimed to detect significant changes in user behavior distribution through statistical tests, drawing inspiration from concept drift detection algorithms in machine learning. However, the applicant further identified the following problems with this scheme: it could only detect whether a change had occurred, but not what that change meant; even after detecting a distribution change, an additional mechanism was needed to determine whether it was noise, a trend, or a shift in the user's experience; it struggled to handle multi-dimensional parallel drift, as users might change their preferences simultaneously at different rates across multiple dimensions, making it difficult for a single detector to cover all of them; and it was disconnected from natural language output, as the detection results were numerical signals that were difficult to directly translate into readable preference understanding updates.

[0022] In summary, applicants have observed that some of the solutions described above treat preference drift as a numerical / statistical problem (using mathematical methods such as weight decay, frequency thresholds, and time windows) rather than a semantic understanding problem. Furthermore, they all employ a single perspective for judgment, facing an irreconcilable contradiction between overfitting and underfitting. They also struggle to distinguish semantically whether preferences have truly changed or if the change is merely a matter of different scenarios. Overfitting, or overreaction, misinterprets occasional behavior as a preference change (e.g., a user occasionally searches for an infrequently used brand, and the system immediately assumes a preference shift). New behaviors too quickly override old preferences, leading users to perceive the system's memory as unstable. Underfitting, or insufficient response, occurs when user preferences have indeed changed, but the system reacts sluggishly and fails to update, still relying on outdated user understanding. Faced with sudden shifts (e.g., moving), the system maintains the old profile for an extended period.

[0023] Based on this, the applicant believes that determining whether a change in behavior is noise or a genuine preference shift essentially requires semantic reasoning of user behavior (for example, "a user starts searching for coffee shop A" in the context of "simultaneous business banquet consumption" may be a scenario expansion rather than a change in taste). This kind of cross-behavioral semantic association reasoning is difficult for statistical methods and hard rules to handle.

[0024] To address these issues, this application proposes a solution. The main idea includes: proposing an adversarial dual-agent debate architecture: an agent with pre-defined changed preferences challenges the other agent, while another agent with unchanged pre-defined preferences defends. Through multiple rounds of structured debate, the true preference state is approximated within the tension of argumentation and questioning. Furthermore, an asymmetric burden of proof mechanism (the challenging agent bears the burden of proof, and if the timeout expires, the conservative stance of the defending agent is assumed) addresses the fundamental contradiction of overfitting and underfitting from a game structure perspective rather than through parameter tuning. The solution will be further detailed below.

[0025] Figure 1This diagram illustrates a user preference analysis and processing method provided in one or more embodiments of this specification. This method can be applied to artificial intelligence (AI) systems that perform user understanding based on user behavior data, such as AI agent dialogue systems (which need to continuously understand user preferences to provide personalized services), intelligent assistant products (which need to remember and track changes in user preferences over a long period), user profiling platforms (which need to dynamically maintain user preference models), or personalized content / service systems (which need to make decisions based on user preferences), etc.

[0026] Figure 1 The process includes the following steps: S102: Obtain user behavior change data.

[0027] For the same user, the corresponding user behavior change data is obtained by comparing the user's behavior data obtained in different time ranges; similarly, updated user behavior change data is obtained based on the user's continued user behavior data in subsequent time ranges.

[0028] Based on user behavior change data over a specific period, the following analysis process is performed to determine whether the user's preferences (which may include one or more) have changed during that period, and if so, how they have changed. Similarly, based on the continued updates to user behavior change data, changes in user preferences can be dynamically tracked periodically or irregularly. This process can be particularly considered for periodic execution. In this case, the user behavior change data could be the user behavior change data corresponding to the current period. For example, assuming a week as a period, the user behavior change data could be the change in a user's behavior data this week compared to the user's behavior data last week (or reference data obtained based on analysis of more historical data).

[0029] For the system, it can analyze or track the user preferences of each user it is interested in extensively, as needed.

[0030] This application does not limit the specific data type of user behavior. For example, it can be user behavior data such as transactions, searches, clicks, browsing, conversations, or location. User behavior data can typically be obtained from the applications used by the user based on user authorization. For example, it can collect user transaction data from e-commerce applications, user payment data or financial product browsing data from payment applications, and so on.

[0031] In one or more embodiments of this specification, differential computation and structured processing may be introduced to more accurately and fairly quantify changes in user behavior and reduce semantic ambiguity, which can support subsequent precise semantic expression.

[0032] Based on this approach, specifically, the process involves obtaining the user's current periodic behavior data; obtaining the user's historical periodic baseline behavior data; and generating user behavior change data for the current period by performing a difference calculation between the current periodic behavior data and the historical periodic baseline behavior data. The historical periodic baseline behavior data can be the user's behavior data from the previous period, or relatively stable and normal user behavior data obtained by comprehensively considering historical data. Therefore, it has good reference value and can be used as a comparison baseline.

[0033] If the raw user behavior data does not meet the expected structured description, it can be transformed into structured data, which facilitates more reliable difference calculations. Structured user behavior change data more clearly highlights and quantifies the degree of change in the behavioral parameters of interest, helps reduce data noise, and facilitates more reliable and efficient semantic understanding and reasoning by subsequent large models.

[0034] S104: Determine the first agent configured to be inclined to affirm that user preferences have changed.

[0035] The intelligent agent described in this application is based on a large model, especially a large language model, to obtain stronger semantic understanding capabilities. Specifically, it is an autonomous intelligent agent that uses a large language model as the core of reasoning and integrates tool invocation, memory management, and autonomous decision-making capabilities.

[0036] In one or more embodiments of this specification, a first agent is configured to be inclined to believe that a user's preference has changed. In this case, it can be assumed that the first agent will take the stance that the preference has changed and perform subsequent processing accordingly, and more actively mine relevant data that supports the change in user preference based on this stance.

[0037] The first agent can be configured to indicate that its preference has changed by inputting corresponding large model cue words (e.g., using cue words to directly indicate the position it should take), also known as the first position for easy distinction, and is configured to be inclined to affirm that the user's preference has changed; or it can be configured to be inclined to affirm that the user's preference has changed through pre-trained bias; and so on.

[0038] S106: The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round.

[0039] In practical applications, changes in user behavior do not necessarily mean changes in user preferences. Nevertheless, due to the bias inherent in the first agent, leveraging its large-scale model semantic understanding and reasoning generation capabilities, it can perform semantic analysis on user behavior change data to generate data supporting its bias. This data is used to challenge another agent (the second agent mentioned later) in a debate, asserting that user preferences have changed. It can also claim specific changes and provide corresponding evidence. Therefore, such data is called user preference change argument data, which may contain one or more arguments demonstrating that user preferences have changed.

[0040] The reason "this round" is mentioned in step S106 is because there may be a next round or even more rounds, which will be explained later. This is part of the periodic execution. Figure 1 In the case of a process, these related rounds can belong to the same cycle, and then one or more rounds may be executed in other cycles as well.

[0041] In addition to user behavior change data, more data can be provided to the first agent to enable it to conduct more comprehensive and accurate semantic analysis and generate more persuasive or insightful arguments about changes in user preferences. For example, providing user behavior data for the current period or previously identified user preferences.

[0042] S108: Determines that the second agent is configured to tend to deny changes in user preferences.

[0043] In one or more embodiments of this specification, the second intelligent agent is configured to tend to deny that the user's preference has changed. In this case, it can be considered that the second intelligent agent will hold the position that the preference has not changed and will conduct subsequent processing accordingly. Based on this position, it will more actively mine relevant data that supports the user's preference not having changed. The first intelligent agent essentially challenges the second intelligent agent, while the second intelligent agent responds and argues based on an adversarial position, trying to question the claims and evidence of the first intelligent agent in order to protect the previously identified user preference. Thus, the two engage in one or more rounds of debate to gradually reach a partial or complete consensus, which helps to approach the true preference state. Therefore, the first intelligent agent can also be figuratively called the challenger intelligent agent, and the second intelligent agent can be called the guardian intelligent agent.

[0044] The second agent can be instructed to hold an unchanged stance against the first stance by inputting corresponding large model cue words (e.g., using cue words to directly indicate the stance it should take), which can also be called the second stance, and is configured to tend to deny that the user's preference has changed; or, it can be configured to tend to deny that the user's preference has changed through pre-trained bias; and so on.

[0045] S110: The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data for the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round.

[0046] The second agent, leveraging the semantic understanding and reasoning capabilities of a large model, performs semantic analysis on data regarding user preference changes. Because the second agent takes an adversarial stance, it will by default question the data on user preference changes to safeguard the stability of user preferences. Of course, the second agent's questioning must also be based on evidence or reasoning, such as insufficient evidence, logical problems, or contradictory data. In such cases, the second agent can generate corresponding user preference change rebuttal data to counter or even refute the first agent's arguments. The first agent can then either abandon its arguments or continue to defend them in the next round, offering stronger evidence and further arguments. Conversely, if the second agent believes there is little room for questioning and that unreasonable questioning is unnecessary, it can generate corresponding user preference change consensus data, indicating at least partial acceptance of the first agent's arguments. For example, it might agree (full acceptance) or acknowledge obvious problems but require further observation or other downgrades (partial acceptance).

[0047] Since the user preference change argument data contains one or more arguments, the second agent can perform the above processing for each argument separately. For example, in the same round, it is possible to generate user preference change defense data for some arguments, and user preference change consensus data for other arguments.

[0048] In addition to data on arguments regarding changes in user preferences, more data can be provided to the second agent to enable it to conduct more comprehensive and accurate semantic analysis, generating more persuasive or insightful data on objections to changes in user preferences and / or data on consensus reached on changes in user preferences. For example, providing reference data on user preferences (including one or more user preferences, which can be considered as the user's previous preferences and can be used as a reference for analyzing changes), data on changes in user behavior, etc.

[0049] In one or more embodiments of this specification, if a complete consensus is not reached after the first round of debate between the first intelligent agent and the second intelligent agent, the first intelligent agent can generate supplementary user preference change argument data in the next round in response to the user preference change defense data generated by the second intelligent agent, in order to continue challenging the second intelligent agent. The second intelligent agent will then continue to defend or reach a consensus in the next round accordingly. Alternatively, the first intelligent agent can also abandon the arguments defended by the second intelligent agent, thereby also achieving a consensus.

[0050] S112: Based on one or more rounds of processing by the first agent and the second agent, determine the changes in user preferences.

[0051] As described above, similarly, the first and second intelligent agents may engage in one or more rounds of debate. If a complete consensus is reached, changes in user preferences can be reliably determined. For example, it can be determined whether user preferences corresponding to different arguments have changed, and if so, how exactly they have changed. Furthermore, the user preference reference data can be updated accordingly (e.g., adding or removing user preferences, modifying the state or content of user preferences), reflecting the user preferences currently recognized by both agents, and can serve as a reference for the future.

[0052] If a complete consensus cannot be reached after multiple rounds of debate and interaction, the debate and interaction can be terminated if the preset convergence conditions are met.

[0053] Taking the case of periodic processing as an example, for instance, based on a set round threshold or time threshold, it can be determined whether the processing of multiple rounds by the first and second agents in the current period meets the threshold convergence condition (i.e., whether the number of rounds or timeouts have been exceeded). If it meets the condition, then in the current period, no further round of processing will be carried out to wait for the next period, and the user preference reference data will be updated accordingly for the arguments that have reached a consensus. It can also be determined whether there are any controversial arguments in the user preference change argument data generated by the first agent. If so, the controversial arguments can be temporarily processed according to the inclination of the second agent, and can be retained or abandoned.

[0054] Furthermore, in the next cycle, user preference analysis is performed again based on the user behavior change data for that cycle, in order to redetermine the changes in user preferences for the next cycle. If any points of contention from the previous cycle are retained, they can be further processed by the first and second intelligent agents based on the new user behavior change data.

[0055] The dynamically determined user preferences tracked can support relevant business parties, such as providing them to downstream consumers for purposes like recommendation information.

[0056] pass Figure 1 The method described in this application utilizes an adversarial dual-agent active debate architecture. A first agent, configured to affirm that user preferences have changed, acts as the challenger. Based on current user behavior change data, as well as the semantic understanding and content generation capabilities of a large-scale model, the first agent initiates arguments regarding the change in user preferences. A second agent, configured to deny the change in user preferences, acts as the defender, responding to the first agent's challenge. Based on the semantic understanding and content generation capabilities of the large-scale model, at least partial rebuttals and / or at least partial consensus are reached. This process can continue for one or more rounds, allowing for a closer examination of the true state of user preferences through the tension between argumentation and questioning. This provides a better understanding of the current user preferences of the corresponding user. Furthermore, newly generated user behavior change data can be used to periodically or irregularly repeat the same process to achieve dynamic adversarial analysis over a longer timeframe, thus contributing to a more reliable and accurate tracking of changes in user preferences.

[0057] based on Figure 1 In addition to the method described herein, this specification also provides some specific implementation schemes and extension schemes of this method, which will be further explained below.

[0058] In one or more embodiments of this specification, an asymmetric burden of proof mechanism is employed. The first agent, acting as the challenger, is primarily responsible for providing evidence. The first agent needs to present sufficient evidence to try and persuade the second agent, acting as the defender. The second agent can refute the first agent's arguments simply by pointing out flaws in the evidence. Furthermore, for arguments that remain controversial after multiple rounds of debate, the second agent's conservative stance can be implicitly adopted. This mechanism, from a game theory perspective, ensures the system's conservative bias, erring on the side of overlooking potential real changes (which can be captured later). It also helps avoid noise contamination of currently identified user preferences, effectively reducing or eliminating the need for manually set threshold parameters.

[0059] Based on this, a specific scheme for debate interaction is proposed as an example.

[0060] For the first intelligent agent, semantic analysis can be performed on user behavior change data, and based on this, the following actions can be taken: Identify one or more change signals in the user behavior change data, such as new behaviors, significant frequency changes, or cross-domain correlations; generate argument data reflecting one or more arguments for each change signal, and use these argument data as the user preference change argument data for this round, or further integrate and process them to obtain user preference change argument data. The arguments mentioned here can include claims and corresponding evidence, primarily derived from user behavior change data or user behavior data. Claims, for example, raise questions about whether user preferences have changed or what changes have occurred. Evidence data can include user behavior data supporting the corresponding claim or reprocessed data derived from it (e.g., interim reasoning results, summaries, or comparison tables).

[0061] It is important to emphasize that if a consensus cannot be reached in one round, the first agent can further analyze the data to supplement stronger evidence to continue supporting its argument, or it can revise its argument. The supplementation of evidence can be progressively strengthened in each round. For example, in the first round, some original text or appropriate summaries can be extracted from user behavior change data or user behavior data as evidence. In the second round, cross-domain correlation analysis tools can be used to more deeply explore the semantic relationships between data from different domains, such as exploring the semantic relationships between user behavior and scenarios, or the hidden semantic relationships between different types of user behavior data, and so on.

[0062] In one or more embodiments of this specification, the user preference change argument data in this round or the user preference change argument data subsequently supplemented can be generated in the following manner: acquiring cross-domain user behavior data involving multiple different dimension domains; performing semantic association analysis on the cross-domain user behavior data using a cross-domain association analysis tool based on a large model to generate cross-domain association analysis data to represent the cross-domain semantic association between user behavior data in different dimension domains, thereby enabling a more comprehensive and accurate joint reflection of the true semantics of user behavior; and generating corresponding user preference change argument data based on the cross-domain association analysis data.

[0063] Cross-domain correlation analysis tools can operate independently of the first intelligent agent, be invoked by the first intelligent agent, or be integrated into the first intelligent agent. These tools can be pre-configured with a semantic analysis logic framework. By structurally reorganizing the acquired multi-dimensional, heterogeneous cross-domain user behavior data, they transform it into contextual input that a large model can understand. Then, leveraging the deep semantic capabilities of the large model in commonsense reasoning, implicit relation mining, and contextual understanding, they perform joint probability assessments and logical deductions on originally isolated behavioral dimension domains (such as transactions, search, browsing, spatiotemporal data, etc.) to uncover semantic correlations between data from different domains, thereby providing more arguments or evidence to support changes in user preferences.

[0064] More specifically, cross-domain association analysis tools can exemplarily employ cue word engineering to implement the aforementioned reasoning logic. In this case, cue word templates can be pre-defined within the tool. These templates, through constraints such as role anchoring, semantic association types, inference link decomposition, and output format, guide the large language model to perform inference and mine semantic association information such as time association and scene association.

[0065] For the second intelligent agent, user preference reference data can be acquired. The second intelligent agent then performs semantic analysis on the user preference reference data, user behavior change data, and user preference change argument data, and accordingly performs: evidence review of each argument reflected in the user preference change argument data; generating corresponding user preference change defense data for this round based on arguments rejected through evidence review; and / or generating corresponding user preference change consensus data for this round based on arguments accepted through evidence review. The evidence review methods are diverse, such as reviewing whether the evidence is sufficient, whether the evidence conforms to the facts, or whether the preferences reflected by the evidence are correct, etc.

[0066] By analogy, multi-round debate-driven stress testing of evidence can be achieved. Through multiple rounds of interactive debate, in-depth stress testing of changing evidence is realized. In each round, the second agent, acting as the guardian, points out specific flaws in the evidence (e.g., insufficient sample size, failure to rule out alternative interpretations, lack of cross-domain corroboration), while the first agent, acting as the challenger, is forced to use tools such as cross-domain correlation analysis to supplement stronger evidence or revise arguments. This iterative process of argumentation, questioning, and supplementary evidence forces the system to continuously deepen semantic analysis, uncovering deep behavioral connections that are difficult to reach in a single inference.

[0067] Furthermore, this application identifies and summarizes several specific types of preference changes, and, based on arguments accepted through evidence review, details how user preferences should be updated across various types of user preference changes. Here, we use data based on consensus regarding user preference changes as an example to indicate the update strategy.

[0068] These specific types of preference changes include: gradual drift, event-driven mutation, and scene switching.

[0069] Gradual drift refers to a slow shift in user preferences, such as gradually shifting from a preference for fast food to a preference for healthy eating. If the corresponding user preference change type is gradual drift, then the consensus data on user preference changes in this round is generated accordingly to indicate whether to retain the old preferences and add the new preferences being observed. In this way, it can adapt to gradual drift, correspondingly and gently switch between old and new preferences, and facilitates rollback.

[0070] Event-driven mutation types refer to drastic changes in user preference patterns caused by events (such as changing jobs, moving, or getting married). If the corresponding user preference change type is an event-driven mutation type, then consensus data on user preference changes for this round is generated to indicate the addition of new preferences as the current preference. This allows for timely responses to important events, aligning with user intuition.

[0071] Scene switching type refers to users exhibiting different preferences in different scenarios (e.g., weekdays and weekends are different scenarios, business trips and staying at home are different scenarios), while the underlying preferences remain unchanged. If the corresponding user preference change type is a scene switching type, then the corresponding user preference change consensus data for this round is generated to indicate whether to retain old preferences, add new preferences, and label the applicable scenarios. In this way, it can adapt to various different scenarios, which helps to improve user immersion.

[0072] Additionally, you can define noise types that might interfere with the judgment of preference changes. These can be ignored to keep the current user preferences unchanged.

[0073] To meet the needs of refined analysis, more selectable states can be created for user preferences, allowing them to be updated to reflect the current situation. Based on this, user preference reference data can include one or more user preferences, along with corresponding state labels for each preference. This allows for state updates based on whether the user preference has changed. For example, at least one state label can represent one of the following states: stable, under observation, changing, controversial, or historical. In addition to these, more states can be created as needed, or existing states can be further subdivided.

[0074] The current type of preference change can be analyzed and determined by the first and / or second intelligent agents. Through the detailed classification of the above types and states, and the corresponding differentiated and refined update strategies, user preferences can be maintained, updated, and tracked more accurately and specifically.

[0075] In one or more embodiments of this specification, it is optional to consider the first agent making more confusing challenges to the second agent to further increase the tension of the debate. Based on this idea, a user preference change argument data processing scheme is proposed, see [link to relevant documentation]. Figure 2 , Figure 2 This is a flowchart illustrating the data processing scheme for the argument of changing user preferences.

[0076] Figure 2 The process includes the following steps: S202: Construct a change signal that does not conform to the facts, and generate argument data for it that includes the claim that does not conform to the facts and the corresponding evidence.

[0077] In addition to extracting change signals from user behavior change data, the first intelligent agent will also actively construct change signals that do not conform to the facts (for example, do not conform to user behavior change data), thereby deliberately constructing erroneous argument data so as to mix it with normal argument data.

[0078] S204: Determine the user preference change argument data for this round based on the argument data corresponding to the identified change signal (which the first agent considers to be consistent with the facts) and the argument data corresponding to the constructed change signal that does not conform to the facts.

[0079] The first intelligent agent clearly understands that the data on changes in user preferences contains data on changes that it has constructed that do not conform to the facts. Under this premise, it still challenges the second intelligent agent to force the second intelligent agent to conduct more in-depth analysis and identification.

[0080] S206: If the second intelligent agent fails to provide sufficient rebuttal to the argument data corresponding to the factually incorrect change signal in the user preference change argument data, it actively generates corresponding corrective data, which is included in the supplementary user preference change argument data.

[0081] The second agent may fail to detect problems with the data in its arguments that do not conform to the facts, or it may find that the objections are not strong enough. In such cases, the second agent can make detailed corrections, thereby guiding it to better protect the data and helping to eliminate blind spots in its understanding.

[0082] Conversely, if the second agent fails to adequately refute the inaccurate argument data, the first agent can abandon the argument. In this case, the first agent can further increase the construction strength of the inaccurate change signals and / or increase the generation strength of the corresponding inaccurate argument data, making it more difficult for the second agent to correctly identify these data, thereby attempting to continue the challenge.

[0083] In one or more embodiments of this specification, user preference reference data may include user preference snapshots. A user preference snapshot is a natural language description of a user's preferences at a specified point in time, which can more accurately and comprehensively reflect a user's preferences at a specific moment on a timeline, allowing for dynamic and flexible retrieval of the required reference when tracking user preferences. User preference snapshots of the same user at one or more different times can be persistently stored for retrieval and updating.

[0084] When determining changes in user preferences based on one or more rounds of processing by the first and second agents, the user preference snapshot can be updated accordingly for reference in the next cycle.

[0085] In addition, to improve the independent thinking of the first intelligent agent, access rights control or business logic control can be used to prevent the first intelligent agent from using user preference snapshots, especially historical user preference snapshots, during the user preference analysis and processing process. Instead, it can make independent inferences based on behavioral data, which is closer to the user's original data and avoids being anchored by data that is more inclined to conclusions, such as user preference snapshots, and thus affects its ability to challenge.

[0086] Based on the foregoing description, and for ease of understanding and implementation, and more intuitively, one or more embodiments of this specification also provide... Figure 1 One specific implementation scheme of the method is shown in [reference]. Figure 3 , Figure 3 This is a schematic diagram of the system architecture for this specific implementation plan, combined with... Figure 3 The specific implementation plan is described in detail.

[0087] Figure 3 The system mainly includes a behavior difference calculation module, a challenger agent (as the first agent mentioned above), a guardian agent (as the second agent mentioned above), and an adversarial debate engine. The preference update process is modeled as a structured debate between two autonomous agents holding opposing pre-set positions.

[0088] The Challenger agent uses a large language model as its reasoning core and is equipped with tools for acquiring behavioral data and cross-domain correlation analysis. Its pre-set stance is "preferences have changed," so it tends to affirm that preferences have changed. It is responsible for detecting change signals from behavioral data, constructing arguments, and presenting evidence in debates to try and persuade the Guardian agent.

[0089] The Guardian agent, using a large language model as its reasoning core, is equipped with tools for accessing and updating historical snapshots. Its default stance is "preferences have not changed," thus it tends to deny that preferences have changed. It is responsible for safeguarding the stability of existing user preference snapshots and questioning flaws in the evidence of change presented by the Challenger agent.

[0090] Two agents approach a true preference state through multiple rounds of debate: the challenger agent proposes a change argument and provides evidence, the guardian agent questions and points out insufficient evidence, and the challenger agent supplements stronger evidence or abandons its argument until both parties reach a consensus or trigger a timeout convergence. An asymmetric burden of proof mechanism is designed: the challenger agent bears the burden of proof, while the guardian agent only needs to raise questions to refute the argument. When the debate times out and no consensus is reached, the conservative position of the guardian agent is adopted by default, ensuring the system's conservative bias.

[0091] I. About the Behavioral Difference Calculation Module.

[0092] Objective: To transform raw behavioral data into structured descriptions of changes and reduce input noise in large language models.

[0093] Input: User behavior data for the current period (e.g., the most recent week) + baseline of behavior statistics for the previous period.

[0094] The processing logic includes: new behavior recognition, which is a behavior that has never appeared or has rarely appeared in history; reduced behavior recognition, which is a behavior that has been high in history but has decreased significantly in the current period; continuous behavior recognition, which is a behavior that is consistent with the historical baseline and has no significant changes; and frequency change calculation, which calculates the frequency change rate of each behavior dimension in the current period compared with the historical baseline.

[0095] Intuitively, an example of the output is as follows: User behavior differential data (as the aforementioned user behavior change data) is calculated by comparing the current period with a historical baseline, including new behaviors and decreased behaviors. New behaviors are assumed to include: 3 purchases of Brand B (in the most recent week, compared to 0 in the previous 90 days); 5 browsing of Category X (a new category). Decreased behaviors are assumed to include: Brand A search frequency decreasing from 15 times / month to 8 times / month (a 47% decrease); Category Y purchase frequency decreasing from 18 times / month to 6 times / month (a 67% decrease). Additionally, there are ongoing behaviors, including: no change in commuting patterns and continued exercise habits.

[0096] This module eliminates the need for large language model calls, which helps reduce costs.

[0097] II. On the Challenger Agent.

[0098] The challenger agent's pre-defined stance is "preferences have changed." It proactively detects change signals from differential user behavior data, constructs one or more arguments that preferences have changed and should be updated, and assumes the burden of proof in the debate. The challenger agent does not access historical snapshots of user preferences, which facilitates independent reasoning.

[0099] An example configuration for the challenger agent is as follows: The core of the reasoning is the large language model, which is responsible for discovering change signals from behavioral data and constructing arguments.

[0100] Toolset: Behavioral data query tool for obtaining user behavior data for a specified period; Behavioral aggregation tool for denoising and aggregating raw behaviors; Cross-domain association analysis tool for analyzing semantic relationships between different behavioral dimensions.

[0101] Access restrictions: Users are not allowed to read historical user preference snapshots; this is implemented through architectural isolation to ensure independence of stance.

[0102] Input: User behavior differential data, current period user behavior data (autonomously acquired by the agent through tools).

[0103] Examples of the structure of large model prompts (Prompts) input to the challenger agent include: "You are a user preference change detection expert. Your stance is: user preferences may have changed."

[0104] Your task is to identify signals of change from behavioral data and to develop arguments for each signal.

[0105] Require: (1) Identify change signals in behavioral data (new behaviors, significant changes in frequency, cross-domain associations) (2) Build an argument for each changing signal, explaining what type of preference change it represents: Gradual drift type: New behaviors continue to increase, while old behaviors decrease; Event-driven mutation type: simultaneous multi-dimensional upheaval; Scene switching type: New and old behaviors alternate in different scenes.

[0106] (3) Each argument contains: Claim: What changes have occurred in preferences? Evidence: Behavioral data supporting this claim; Self-assessment of evidence strength: Strong (multiple times, cross-domain corroboration) / Medium (multiple times but single dimension) / Weak (only 1~2 times).

[0107] (4) For arguments with weak evidence, proactively explain their weaknesses, but still raise them.

[0108] User behavior differential data, obtained from: {behavioral_diff}; User behavior data, obtained from: {behavior_data}.

[0109] Output: A set of structured argument data on preference changes, each argument including the claim, evidence, type of change, and self-assessment of the strength of evidence.

[0110] For cross-domain association analysis tools, a large model is used as the inference engine, combined with a cross-domain association analysis prompt. The user's cross-domain behavior is input, and the analysis conclusions are output. For example, the following cross-domain association analysis prompt can be used: # Role You are a cross-domain user behavior correlation analysis engine. Your task is to cross-analyze user behavior data scattered across different behavioral dimensions (domains) and uncover the implicit semantic relationships between them.

[0111] # Core Concepts A single behavioral dimension cannot reveal the complete semantics of a behavior. Only by intersecting behaviors across multiple dimensions (such as transactions, searches, browsing, location, time, context, and social interactions) can the true semantic relationships be identified.

[0112] # Input You will receive a structured piece of user cross-domain behavior data, which typically includes: The current period and the historical baseline are used to differentiate the data. Details for each behavioral dimension (timestamp, frequency, object, scene tag, etc.) # Analysis Requirements Follow these steps to complete the analysis, do not skip any: 1. Domain Identification List all behavioral dimensions (domains) involved in the input data, and label the key signals observed in each domain.

[0113] 2. Cross-domain association mining When searching for semantic relationships between domains, the following categories should be the focus: Time correlation: Whether the behaviors of different domains are highly consistent within the same time window / period / specific date, and provide the consistency rate.

[0114] Scenario association: Whether a behavior in a certain domain is stably bound to a specific scenario (weekday / weekend / business trip / banquet).

[0115] The rise and fall of one domain: Does the rise of one domain coincide with the decline of another, or do the two coexist (meaning the differentiation of scenarios rather than replacement)?

[0116] Co-occurrence patterns: Which behaviors always appear in pairs or groups, and what semantics their combinations point to.

[0117] Anomalies: Are there any outliers that do not conform to the main pattern?

[0118] # Output Format Output according to the following structure: [Cross-domain association analysis conclusions] 1. The behavioral domains involved: Domain 1: Key Signal Domain 2: Key Signals ... 2. Cross-domain association discovery: Relationship ①: <Describe the relationship in one sentence> Association types: Time association / Scene association / One-to-one relationship / Co-occurrence pattern Quantitative basis: <Specific figures such as consistency rate, frequency comparison, and time window overlap> The semantic meaning of the association: <This association describes what the user is doing> Related ②: ... # Constraints The analysis is based solely on the input data and does not fabricate behaviors not provided.

[0119] All quantitative conclusions must be supported by specific numerical data, and vague wording should be avoided. By solidifying and reusing these prompts, the tool can stably output cross-domain correlation analysis data with a unified structure, logical consistency, and traceable evidence across different user instances, effectively supporting the automated generation of arguments for subsequent preference changes.

[0120] III. About the Guardian Intelligence Agent.

[0121] The Guardian agent's default stance is "preferences have not changed." It safeguards the stability of the existing snapshot of user preferences and questions and scrutinizes the evidence for every argument put forward by the Challenger agent. The Guardian agent holds the right of adjudication; if no consensus is reached within the debate timeout, the Guardian agent's conservative stance is adopted by default.

[0122] An example configuration for the Guardian agent is as follows: The core of the reasoning is the large language model, which is responsible for evidence review, rebuttal construction, and final ruling.

[0123] Toolset: User preference snapshot reading tool, used to retrieve stored user preference snapshots; User preference snapshot writing tool, used to persist updated snapshots; Historical baseline comparison tool, used to compare current behavior with historical baselines over a longer period.

[0124] Permission configuration: It has read and write permissions for user preference snapshots, forming a permission asymmetry with the challenger agent.

[0125] Ruling power: When the debate exceeds the time limit or the number of rounds is reached, the Guardian agent's position is taken as the default result.

[0126] Input: Data on the challenger agent's arguments about changes in user preferences, currently stored snapshots of user preferences, and differential data on user behavior.

[0127] Examples of the structure of large model prompts input to the guardian agent include: "You are a user preference stability expert. Your position is: user preferences should remain unchanged until sufficient evidence is available."

[0128] You will receive the challenger agent's arguments for change, as well as current user preference snapshots and differential user behavior data.

[0129] For each argument of the challenger agent, you must complete the following review: Step 1: Review of the sufficiency of evidence.

[0130] Examine each argument for sufficiency of evidence: Is the sample size sufficient? (It needs to be considered whether only 1-2 occurrences are enough to support the conclusion). Have alternative explanations (promotional, random, seasonal, etc.) been ruled out? Is there cross-domain corroboration (a change in a single dimension, or a multi-dimensional linkage)? The degree of conflict with historical preference snapshots (minor adjustments or fundamental changes).

[0131] Step 2: Refute or accept.

[0132] Give a clear ruling on each argument: (1) Rejection: The evidence is insufficient. Please specify which pieces of evidence are missing. (2) Partial acceptance: The core claim is valid but needs to be downgraded. For example, accept that "there are signs of change" but mark it as "under observation" instead of "confirmed". (3) Full acceptance: The evidence is sufficient to accept the claims and classifications of the challenger agent.

[0133] Step 3: Based on the accepted arguments, decide on an updated strategy.

[0134] (1) Gradual drift type: retain the old preferences, add new preferences and mark them "under observation"; (2) Event-driven mutation type: old preferences are marked as "historical preferences", and new preferences are upgraded to current preferences; (3) Scene switching type: old and new preferences coexist, and the applicable scenarios are marked respectively.

[0135] In addition, noise, such as rejected arguments, can be ignored, and existing preferences can be maintained.

[0136] You must provide the basis for your review and ruling for each argument, without skipping any points. For arguments that are rejected, you must clearly point out the specific flaws in the evidence.

[0137] The current user preference snapshot is obtained from: {current_snapshot} (an exemplary data structure for storing the current user preference snapshot); user preference change argument data is obtained from: {challenger_claims}; user behavior difference data is obtained from: {behavioral_diff}.

[0138] Output: A point-by-point ruling for each argument (rejection / partial acceptance / full acceptance), along with a detailed explanation of the flaws when rejected. For accepted arguments, output the update strategy and drift classification.

[0139] IV. On the adversarial debate engine.

[0140] It supports challenger and guardian agents to approach accurate judgments of preference changes through multiple rounds of structured debate, amidst the tension of argumentation and questioning.

[0141] The debate process: First round (first round of debate): The challenger agent proposes arguments for change, accompanied by initial evidence; The Guardian AI reviews each statement, rejecting / partially accepting / fully accepting it.

[0142] Second round (supplementary debate, triggered only when there is a rebuttal): The challenger agent can supplement stronger evidence or abandon rejected arguments, and can use cross-domain correlation analysis tools to obtain new evidence; The Guardian AI agent reviews the supplementary evidence and issues a ruling.

[0143] The third round (assuming a maximum of 3 rounds, this is the final round, triggered only if there is still controversy) Challenger agent, final supplementary argument; The Guardian agent makes the final decision and outputs an updated snapshot of user preferences.

[0144] Convergence conditions (the debate in the current cycle terminates when any one of the conditions is met): (1) Consensus convergence.

[0145] Triggering scenario: The guardian agent fully accepts all arguments, or the challenger agent abandons all rejected arguments; Result: The user preference snapshot was updated according to the consensus reached by both parties.

[0146] (2) Convergence beyond the threshold.

[0147] Triggering scenario: There is still controversy after reaching the maximum number of debate rounds (default 3 rounds); Result: The controversial arguments were handled according to the conservative stance of the Guardian intelligence, either rejected or downgraded to "under observation".

[0148] Debate transcripts and outstanding disputes: The argumentation process and rulings of each round of debate are recorded in a structured manner and stored in a preference snapshot.

[0149] For matters that have been resolved, changes agreed upon by both parties will be directly updated in the preference snapshot.

[0150] For unresolved disputes, arguments that are temporarily rejected during the threshold convergence are marked as "to be verified in the next cycle". In the next cycle of debate, the challenger agent can give priority to these controversial points.

[0151] Output: An updated snapshot of user preferences (described in natural language), including a summary of the debate proceedings and a list of unresolved points of contention. The Guardian agent autonomously performs persistent storage via a snapshot writing tool.

[0152] V. Status label system for user preference snapshots.

[0153] Several optional status labels are defined for user preferences as an example, as shown in the table below:

[0154] An example flow update path for status tags is as follows: Under observation → Stable (Challenger agents provide continuous evidence for 2-3 consecutive cycles, and the Guardian agents fully accept it) During observation → Elimination (The challenger agent will no longer provide evidence for this argument in subsequent cycles). Under observation → Under dispute (debate timed out, no consensus reached) Controversial → Under Observation (The challenger agent provided new evidence in the next cycle) In dispute → Eliminated (the challenger agent abandons the argument in the next cycle) Stable → Changing (The Challenger agent proposes the asymptotic drift argument, which the Guardian agent partially accepts) In transition → Stable (New preferences are fully accepted and confirmed by the Guardian agent) In flux → History (Old preferences are accepted and replaced by the Guardian agent) Stable → Historical (The challenger agent presents sufficient evidence of event-driven mutations, which the guardian agent fully accepts) VI. Regarding the operating cycle and scheduling strategy.

[0155] Examples are shown in the table below:

[0156] VII. An exemplary workflow in a specific scenario.

[0157] Scenario: The user was originally a heavy customer of coffee shop B, but recently started making purchases at coffee shop A.

[0158] First cycle: User behavior differential data: New additions: 2 purchases at coffee shop A; Continuing trend: 12 searches for coffee shop B per month (slightly decreasing).

[0159] Debate process: Round 1: Challenger agent: "Detected new change signal, coffee shop A made 2 purchases."

[0160] Claim: Users' coffee preferences are beginning to shift towards coffee shop A; Evidence: Coffee shop A's customer spending increased from 0 to 2 times; Type of change: Gradual drift (early signal) Self-assessment of evidence strength: Weak (only 2 instances, no cross-domain corroboration). Guardian AI: "Rejected."

[0161] Flaw (1): The sample size is only 2 times, which is not statistically significant; Vulnerability (2): The behavior of coffee shop B did not decrease significantly (only slightly decreased to 12 times / month), which does not conform to the ebb and flow characteristics of gradual drift; Vulnerability (3): No evidence of cross-domain behavior (e.g., no related behavior of coffee shop A in search, browsing, etc.) Ruling: Noise; user preference remains unchanged as "deep user of coffee shop B". Challenger agent: "Accept rejection, abandon the argument, the evidence is indeed insufficient." Consensus can be reached in just one round, and user preferences remain unchanged.

[0162] Second cycle (assuming 2 weeks later): User behavior differential data: Up, 5 visits to coffee shop A (significant increase); Down, 6 searches for coffee shop B per month (50% decrease); New, 3 visits for business banquets.

[0163] Debate process: Round 1: Challenger agent: "Multiple associated change signals have been detected."

[0164] Argument A: Users' coffee preferences are shifting from coffee shop B to coffee shop A; Evidence: Visits to coffee shop A increased to 5 times, while searches for coffee shop B decreased by 50%. Type of change: gradual drift; Self-assessment of the strength of evidence: Medium (there is a trend of one side increasing while the other decreases, but the reason is unclear); Argument B: New business and social scenarios are emerging for users; Evidence: The three instances of business banquets constitute new activity; Change type: Scene switching; Self-assessment of evidence strength: Medium (The newly added behavior is clear, but its connection to coffee preference needs further proof). Guardian agent: "Argument A is rejected, argument B is accepted."

[0165] Flaw in Argument A: Interpreting the rise of coffee shop A and the decline of coffee shop B as a shift in taste preferences is too hasty. The alternative explanation of scene difference needs to be ruled out. Is coffee shop A consumption concentrated in a specific scene (e.g., weekdays)? Is coffee shop B still active in other scenes?

[0166] Argument B Review: Three business banquets constitute a new behavior, which is a clear signal, but is a sample of three sufficient to confirm a persistent preference?

[0167] Ruling: Partially accepted, marked as under observation. Round 2: Challenger agent: "Provide supplementary evidence to address the flaws in argument A."

[0168] Cross-domain association analysis tools revealed that of the five transactions at coffee shop A, four occurred during weekday lunchtimes, with three of these transactions coinciding perfectly with the dates of business banquets; the transactions at coffee shop B still occurred three times on weekends and did not disappear.

[0169] Revised claim: This is not a shift in taste preferences (gradual drift), but rather a change in setting. Coffee shop A is the choice for business and social settings, while coffee shop B remains the choice for personal settings.

[0170] Combined arguments A and B, we can conclude that user coffee preferences vary across different scenarios. Guardian agent: "Fully accept the revised claims."

[0171] Evidence review passed: temporal correlation established (80% match between coffee shop A and banquet); alternative explanations ruled out (coffee shop B remained active on weekends, not a change in taste); sufficient cross-domain corroboration (linkage between coffee consumption and banquet consumption).

[0172] Ruling: Scenario shift, old and new preferences coexist. Coffee shop B (personal / weekend scenario, stable); Coffee shop A (business / social scenario, under observation).

[0173] Verification point: In the next cycle, we will focus on whether the consumption at coffee shop A continues to be linked to business scenarios. This is achieved through two rounds of consensus, and the user preference snapshot is updated accordingly.

[0174] Third cycle (assuming another 2 weeks): User behavior differential data: Continuous, coffee shop A was visited 4 times, all during weekday lunchtimes; Continuous, coffee shop B was visited 3 times, all on weekends; Continuous, business banquets were held 2 times.

[0175] Unresolved disputes from the previous round: None.

[0176] Previous verification point: Whether consumption at coffee shop A is continuously associated with business scenarios.

[0177] Debate process: Round 1: Challenger agent: "Propose arguments based on the validation points of the previous cycle."

[0178] Argument: The association between coffee shop A and business scenarios has been verified across cycles, and coffee shop A (business social scenario) should be upgraded from observation to stability. Evidence: Consumption at coffee shop A for two consecutive periods closely matched weekday / business scenarios; coffee shop B maintained stable activity on weekends; behavioral patterns were consistent across periods and not sporadic.

[0179] Self-assessment of evidence strength: Strong (cross-period verification + consistent validity based on scenario relevance). Guardian agent: "Completely accept."

[0180] Evidence review passed: cross-cycle consistency verification (the scenario switching judgment in the second cycle was confirmed in the third cycle); stable behavior patterns in both scenarios (weekday coffee shop A + weekend coffee shop B). Ruling: Upgrade confirmed. Coffee shop B (personal / weekend setting, stable); Coffee shop A (work / social setting, stable ← upgrade from observation). Consensus is reached in just one round, and the guardian agent accepts it directly without debate, with user preference snapshots updated accordingly.

[0181] The proposed solution in this application has advantages over some solutions previously attempted by the applicant and helps to overcome some disadvantages, as illustrated by example.

[0182] For the full recalculation scheme, it helps to overcome the disadvantage of unstable results.

[0183] This application's scheme uses a guardian agent that safeguards historical preference snapshots as anchors and holds the right of adjudication. While the challenger agent reasons independently based on behavioral data each time (which may fluctuate), its arguments generally undergo one or more rounds of questioning and evidence review by the guardian agent. The user preference snapshot is only updated after the challenger agent presents sufficient evidence to convince the guardian agent. This dual mechanism of asymmetrical burden of proof and defaulting to the guardian agent's position after a timeout ensures that the system output remains highly consistent across multiple runs, unless there is sufficient evidence, verified through debate, that the preferences have indeed changed.

[0184] For simple add-on schemes, this helps overcome the drawback of overly rapid coverage.

[0185] In this application's scheme, preference changes require a complete process: challenger agent presents evidence → guardian agent challenges the change → multiple rounds of debate → consensus or timeout. The guardian agent reviews the sufficiency of evidence for each change argument, pointing out specific flaws such as insufficient sample size, failure to exclude alternative explanations, and lack of cross-domain corroboration. Noise arguments are rejected in the first round, and even if a new preference during gradual drift is accepted, it is only marked as under observation; new and old preferences coexist rather than overlap during scenario transitions. This, from an adversarial mechanism perspective, prevents the problem of new information overwriting old preferences; overwriting requires winning the debate.

[0186] For statistical thresholding schemes, it helps to overcome the shortcomings of semantic blind spots.

[0187] This application's scheme achieves semantic-level evidence stress testing through adversarial debate. The challenger agent uses cross-domain association analysis tools to discover semantic connections between behaviors (e.g., coffee shop A purchase + business dinner = scenario expansion), while the guardian agent questions the reliability of these connections from the perspective of alternative interpretations. The multi-round interaction of the debate forces both parties to continuously deepen semantic reasoning. The challenger agent needs to respond to the loopholes pointed out by the guardian agent, thereby producing a more reliable semantic understanding.

[0188] For concept drift detection schemes, this helps overcome the shortcoming of only detecting but not judging.

[0189] This application's scheme assigns change detection (the challenger agent's responsibility) and change attribution (clarified gradually during the debate) to different agents, completing an end-to-end closed loop from "detection → attribution → decision" in adversarial debate. The challenger agent not only detects whether there are change signals but also proposes a change type judgment; the guardian agent then examines the reasonableness of this judgment from a guardian's perspective. The debate result includes not only the final ruling but also the complete argumentation process and unresolved points of contention.

[0190] Furthermore, the applicant has also developed some alternative solutions that differ from the solutions described in this application, while the solutions described in this application still possess their own advantages. Three alternative solutions are compared below: Alternative Solution 1: Single Agent Judgment Scheme (No Adversarial)

[0191] A single agent simultaneously performs change detection and stability control, completing preference update judgments within a single inference iteration. Differences from this application include: a single agent simultaneously playing both change detection and stability protection roles is prone to cognitive biases, either leaning towards sensitivity (overfitting) or conservatism (underfitting), depending on the wording of the prompts; the lack of adversarial evidence stress testing means the agent's judgments are not challenged or refuted, making it difficult to expose flaws in the evidence; the absence of a mechanism to ensure asymmetric burden of proof means the system's conservatism relies on the tone of the prompts rather than architectural constraints; and the lack of debate records and unresolved dispute mechanisms makes it difficult to accumulate attention to controversial points across cycles.

[0192] Alternative solution two: pipelined dual-agent solution (cooperation rather than confrontation).

[0193] Two agents collaborate in a pipeline manner: the first agent generates preference understanding, and the second agent reviews and decides to update the strategy. Differences from this application include: the two agents do not have opposing pre-existing stances; the second agent is a neutral reviewer rather than a stance guardian, and the review's effectiveness depends on the effect of the prompts; there are no multi-round debates; the second agent makes a one-time judgment on the first agent's output, and the first agent has no opportunity to supplement evidence or revise arguments; there is a lack of burden of proof allocation, making it impossible to guarantee the system's conservative bias through mechanisms; it is more suitable as a lightweight solution with high efficiency but low prudence, applicable to scenarios where preference stability requirements are not high.

[0194] Alternative solution three: event-driven incremental update solution.

[0195] The large language model is only triggered to re-evaluate preferences when a significant change in behavior distribution is detected by a lightweight statistical detector; otherwise, the previous version of preferences remains unchanged. Differences from this application include: reliance on the sensitivity setting of the statistical detector; a threshold that is too high will miss asymptotic drift, while a threshold that is too low will trigger frequent shifts; the statistical detector cannot identify the semantic nature of behavior changes (whether it is noise or a trend), and lacks semantic reasoning capabilities in agent adversarial debate; it is suitable as an auxiliary optimization of this scheme (skipping the debate process when there is no change in behavior), but cannot replace the semantic drift judgment in adversarial debate.

[0196] In summary, the technical effects achievable by the specific solutions described above in this application include: The architecture is designed to resist overfitting. The asymmetric burden of proof and the default conservative stance during timeouts ensure the system's conservative bias through a game-theoretic mechanism, reducing or even eliminating the need for manual tuning of threshold parameters.

[0197] Evidence stress testing. Each change signal undergoes one or more rounds of confrontation between challenger agent argumentation and guardian agent questioning. Flaws in the evidence are exposed and fixed during the debate, producing preference updates that are more reliable than those made by a single agent.

[0198] Debate-driven deep semantic reasoning. Multi-round interactions force the agent to continuously deepen its analysis (e.g., from increased consumption at coffee shop A to the temporal association between coffee shop A and business scenarios), uncovering cross-domain semantic connections that are difficult to reach in a single reasoning session.

[0199] Cross-cycle dispute tracking. Unresolved disputes are recorded and given priority in the next cycle, forming a cross-cycle evidence accumulation mechanism, so that gradual drift will not be ignored for a long time due to insufficient evidence in a single instance.

[0200] Complete interpretability. The debate transcript contains the complete process of both sides' arguments, evidence, challenges, and rulings, tracing the decision-making path of each preference update, facilitating manual review and system debugging.

[0201] Versatility. The solution is not limited to specific data types of user behavior; the adversarial debate mechanism is applicable to many types of user behavior data sources.

[0202] Based on the same idea, one or more embodiments of this specification also provide apparatus and devices corresponding to the above methods, such as... Figure 4 , Figure 5 As shown. The apparatus and equipment are capable of performing the above methods and related alternatives accordingly.

[0203] Figure 4 This specification provides a schematic diagram of the structure of a user preference analysis and processing device according to one or more embodiments, the device comprising: Behavior change data acquisition module 402 acquires user behavior change data; First agent determination module 404 determines a first agent configured to be inclined to affirm that user preferences have changed; The preference change argument generation module 406 performs semantic analysis on the user behavior change data by the first intelligent agent to generate user preference change argument data for this round. The second agent determination module 408 determines a second agent that is configured to tend to deny changes in user preferences. The preference change argument confrontation module 410 performs semantic analysis on the user preference change argument data through the second intelligent agent to generate user preference change defense data and / or user preference change consensus data in the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. The user preference change determination module 412 determines the changes in user preferences based on one or more rounds of processing by the first agent and the second agent.

[0204] Optionally, the behavior change data acquisition module 402 acquires the user's current periodic behavior data, which is described in a structured manner. Obtain the historical periodic baseline behavior data of the user's structured description; By performing differential calculations adapted to structured descriptions on the current period behavior data and the historical period baseline behavior data, user behavior change data for the user in the current period is generated. After determining the changes in user preferences, the device performs user preference analysis again based on the user behavior change data for the next period, so as to redetermine the changes in user preferences for the next period.

[0205] Optionally, the first agent is configured to indicate its first stance by inputting corresponding large model cue words, and is inclined to affirm that user preferences have changed; and / or, The second agent is configured to oppose the first stance by inputting corresponding large model cue words, and is thus configured to deny that the user's preference has changed.

[0206] Optionally, the preference change argument generation module 406 performs semantic analysis on the user behavior change data through the first intelligent agent and executes accordingly: Identify one or more change signals in the user behavior change data; For each of the aforementioned change signals, argument data containing claims and corresponding evidence is generated, wherein the claims propose what changes have occurred in user preferences, and the evidence data contains user behavior data that supports the corresponding claims; Based on the data of each argument, determine the data of the arguments regarding changes in user preferences in this round.

[0207] Optionally, the preference change argument generation module 406 constructs a change signal that does not conform to the facts, and generates argument data for it containing the claim that does not conform to the facts and the corresponding evidence; Based on the argument data corresponding to the identified change signals and the argument data corresponding to the constructed change signals that do not conform to the facts, determine the argument data for the current round of user preference changes; The preference change argument generation module 406 proactively generates corresponding correction data, which is included in the supplementary user preference change argument data, when the second agent fails to provide sufficient rebuttal to the argument data corresponding to the factually inconsistent change signal in the user preference change argument data.

[0208] Optionally, the current round of user preference change argument data or the supplementary user preference change argument data is generated in the following manner: Acquire cross-domain user behavior data involving multiple different dimensions and domains; Using a cross-domain association analysis tool based on a large model, semantic association analysis is performed on the cross-domain user behavior data to generate cross-domain association analysis data. Based on the cross-domain correlation analysis data, corresponding data on user preference changes are generated.

[0209] Optionally, the preference change argument countermeasure module 410 acquires user preference reference data; The second intelligent agent performs semantic analysis on the user preference reference data, the user behavior change data, and the user preference change argument data, and executes accordingly: The evidence for each argument reflected in the data on changes in user preferences is examined. Based on the arguments rejected by the evidence review ruling, corresponding user preference change defense data for this round is generated; and / or, based on the arguments accepted by the evidence review ruling, corresponding user preference change consensus data for this round is generated.

[0210] Optionally, the preference change argument countermeasure module 410 determines the type of user preference change corresponding to the argument accepted by the evidence review ruling among multiple different types of user preference change, and executes accordingly: If the corresponding user preference change type is a gradual drift type, then the consensus data for this round of user preference changes is generated accordingly to indicate the retention of old preferences and the addition of new preferences under observation; or, If the corresponding user preference change type is an event-driven mutation type, then the consensus data for this round of user preference change is generated accordingly to indicate that a new preference has been added as the current preference; or, If the corresponding user preference change type is a scenario switching type, then the consensus data of the current round of user preference change is generated accordingly to indicate the retention of old preferences, the addition of new preferences, and the marking of applicable scenarios.

[0211] Optionally, the user preference reference data includes a user preference snapshot, which is a natural language description of user preferences at a specified time. The user preference change determination module 412 updates the user preference snapshot accordingly based on the changes in user preferences.

[0212] Optionally, access rights control or business logic control can be used to prevent the first intelligent agent from using the user preference snapshot during the user preference analysis and processing.

[0213] Optionally, the user preference reference data includes one or more user preferences and a status label corresponding to each user preference, so as to update the status accordingly based on whether the user preference has changed. At least one of the aforementioned state labels represents one of the following states: Stable, under observation, changing, controversial, historical.

[0214] Optionally, the user preference change determination module 412 determines, based on a set round threshold or time threshold, whether the processing of the first agent and the second agent in multiple rounds of the current period meets the threshold convergence condition. If the conditions are met, then in the current cycle, the next round of processing will not be carried out to wait for the next cycle, and it will be determined whether the user preference change argument data generated by the first agent contains any disputed arguments; If it exists, then according to the inclination of the second agent, the arguments in the dispute are temporarily processed and reserved; In the next cycle, the device, through the first and second agents, continues to process the arguments in dispute based on new user behavior change data.

[0215] Figure 5 This is a schematic diagram of the structure of a user preference analysis and processing device provided in one or more embodiments of this specification. The user preference analysis and processing device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform: Obtain data on changes in user behavior; The first agent is determined to be configured to be inclined to believe that user preferences have changed. The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round. The second agent is configured to tend to deny changes in user preferences. The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data in the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. Based on one or more rounds of processing by the first agent and the second agent, the changes in user preferences are determined.

[0216] Based on the same idea, one or more embodiments of this specification also provide a non-volatile computer storage medium storing computer-executable instructions configured as follows: Obtain data on changes in user behavior; The first agent is determined to be configured to be inclined to believe that user preferences have changed. The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round. The second agent is configured to tend to deny changes in user preferences. The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data for this round. The user preference change defense data is used by the first intelligent agent to generate supplementary user preference change argument data or abandon arguments in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. Based on one or more rounds of processing by the first agent and the second agent, the changes in user preferences are determined.

[0217] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0218] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0219] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0220] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0222] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0223] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0224] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0225] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0226] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0227] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0228] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A user preference analysis and processing method, comprising: Obtain data on changes in user behavior; The first agent is determined to be configured to be inclined to believe that user preferences have changed. The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round. The second agent is configured to tend to deny changes in user preferences. The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data in the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. Based on one or more rounds of processing by the first agent and the second agent, the changes in user preferences are determined.

2. The method as described in claim 1, wherein obtaining user behavior change data specifically includes: Obtain the user's current periodic behavior data; Obtain the user's historical baseline behavior data; By performing differential calculations on the current period behavior data and the historical period baseline behavior data, structured user behavior change data for the current period is generated. After determining the changes in user preferences, the method further includes: Based on the user behavior change data for the next period, user preference analysis is performed again to redetermine the changes in user preferences for the next period.

3. The method of claim 1, wherein the first agent is configured to tend to affirm that user preferences have changed by inputting corresponding large model cue words to indicate its first stance; and / or, The second agent is configured to oppose the first stance by inputting corresponding large model cue words, and is thus configured to deny that the user's preference has changed.

4. The method as described in claim 1, wherein the step of performing semantic analysis on the user behavior change data through the first intelligent agent to generate user preference change argument data for this round specifically includes: The first intelligent agent performs semantic analysis on the user behavior change data and executes accordingly: Identify one or more change signals in the user behavior change data; For each of the aforementioned change signals, argument data containing claims and corresponding evidence is generated, wherein the claims propose what changes have occurred in user preferences, and the evidence data contains user behavior data that supports the corresponding claims; Based on the data of each argument, determine the data of the arguments regarding changes in user preferences in this round.

5. The method as described in claim 4, wherein determining the user preference change argument data for this round based on each of the argument data specifically includes: Construct false change signals and generate argument data containing false claims and corresponding evidence; Based on the argument data corresponding to the identified change signals and the argument data corresponding to the constructed change signals that do not conform to the facts, determine the argument data for the current round of user preference changes; The method further includes: If the second agent fails to adequately refute the argument data corresponding to the factually incorrect change signals in the user preference change argument data, it proactively generates corresponding corrective data, which is included in the supplementary user preference change argument data.

6. The method as described in claim 4, wherein the current round of user preference change argument data or the supplementary user preference change argument data is generated in the following manner: Acquire cross-domain user behavior data involving multiple different dimensions and domains; Using a cross-domain association analysis tool based on a large model, semantic association analysis is performed on the cross-domain user behavior data to generate cross-domain association analysis data. Based on the cross-domain correlation analysis data, corresponding data on user preference changes are generated.

7. The method as described in claim 1, wherein the step of performing semantic analysis on the user preference change argument data through the second intelligent agent to generate the user preference change rebuttal data and / or user preference change consensus data for this round specifically includes: Obtain user preference reference data; The second intelligent agent performs semantic analysis on the user preference reference data, the user behavior change data, and the user preference change argument data, and executes accordingly: The evidence for each argument reflected in the data on changes in user preferences is examined. Based on the arguments rejected by the evidence review ruling, corresponding user preference change defense data for this round is generated; and / or, based on the arguments accepted by the evidence review ruling, corresponding user preference change consensus data for this round is generated.

8. The method of claim 7, wherein generating the consensus data on user preference changes in this round based on the arguments accepted by the evidence review decision, specifically includes: Determine the type of user preference change corresponding to the arguments accepted through the aforementioned evidence review ruling among various different types of user preference changes, and execute accordingly: If the corresponding user preference change type is a gradual drift type, then the consensus data of the current round of user preference change is generated accordingly to indicate that the old preferences are retained and the new preferences under observation are added. or, If the corresponding user preference change type is an event-driven mutation type, then the consensus data of the current round of user preference change is generated accordingly to indicate that a new preference has been added as the current preference; or, If the corresponding user preference change type is a scenario switching type, then the consensus data of the current round of user preference change is generated accordingly to indicate the retention of old preferences, the addition of new preferences, and the marking of applicable scenarios.

9. The method of claim 7, wherein the user preference reference data includes a user preference snapshot, the user preference snapshot being a natural language description of user preferences at a specified time; The step of determining changes in user preferences based on one or more rounds of processing by the first agent and the second agent further includes: The user preference snapshot is updated accordingly based on changes in user preferences.

10. The method of claim 1 or 9, further comprising: By controlling access permissions or business logic, the first intelligent agent is prevented from using historical user preference snapshots during the user preference analysis and processing.

11. The method according to any one of claims 7 to 9, wherein the user preference reference data includes one or more user preferences and a status label corresponding to each user preference, so as to update the status accordingly based on whether the user preference has changed; At least one of the aforementioned state labels represents one of the following states: Stable, under observation, changing, controversial, historical.

12. The method as described in claim 1, wherein determining the changes in user preferences based on multiple rounds of processing by the first agent and the second agent specifically includes: Based on the set round threshold or time threshold, determine whether the processing of the first agent and the second agent in multiple rounds of the current cycle meets the convergence condition of exceeding the threshold. If the conditions are met, then in the current cycle, the next round of processing will not be carried out to wait for the next cycle, and it will be determined whether the user preference change argument data generated by the first agent contains any disputed arguments; If it exists, then according to the inclination of the second agent, the arguments in the dispute are temporarily processed and reserved; The method further includes: In the next cycle, the arguments in dispute will continue to be processed by the first and second agents based on new user behavior change data.

13. A user preference analysis and processing device, comprising: The behavior change data acquisition module acquires user behavior change data. The first agent determination module determines the first agent that is configured to be inclined to affirm that the user's preferences have changed. The preference change argument generation module performs semantic analysis on the user behavior change data by the first intelligent agent to generate user preference change argument data for this round. The second agent determination module determines the second agent that is configured to tend to deny changes in user preferences. The preference change argument adversarial module performs semantic analysis on the user preference change argument data through the second intelligent agent to generate user preference change defense data and / or user preference change consensus data for the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. The user preference change determination module determines the changes in user preferences based on one or more rounds of processing by the first agent and the second agent.

14. The apparatus of claim 13, wherein the behavior change data acquisition module acquires the user's current periodic behavior data; Obtain the user's historical baseline behavior data; By performing differential calculations on the current period behavior data and the historical period baseline behavior data, structured user behavior change data for the current period is generated. After determining the changes in user preferences, the device performs user preference analysis again based on the user behavior change data for the next period, so as to redetermine the changes in user preferences for the next period.

15. The apparatus of claim 13, wherein the first agent is configured to tend to affirm that user preferences have changed by inputting a corresponding large model cue word to indicate its first stance; and / or, The second agent is configured to oppose the first stance by inputting corresponding large model cue words, and is thus configured to deny that the user's preference has changed.

16. The apparatus of claim 13, wherein the preference change argument generation module performs semantic analysis on the user behavior change data by the first intelligent agent, and executes accordingly: Identify one or more change signals in the user behavior change data; For each of the aforementioned change signals, argument data containing claims and corresponding evidence is generated, wherein... The claim outlines what changes have occurred in user preferences, and the evidence data includes user behavior data that supports the corresponding claim. Based on the data of each argument, determine the data of the arguments regarding changes in user preferences in this round.

17. The apparatus of claim 16, wherein the preference change argument generation module constructs a change signal that does not conform to the facts, and generates argument data for it containing the claim that does not conform to the facts and the corresponding evidence; Based on the argument data corresponding to the identified change signals and the argument data corresponding to the constructed change signals that do not conform to the facts, determine the argument data for the current round of user preference changes; The preference change argument generation module proactively generates corresponding corrective data, which is included in the supplementary user preference change argument data, if the second agent fails to provide sufficient rebuttal to the argument data corresponding to the factually inconsistent change signals in the user preference change argument data.

18. The apparatus of claim 16, wherein the current round of user preference change argument data or the supplementary user preference change argument data is generated in the following manner: Acquire cross-domain user behavior data involving multiple different dimensions and domains; Using a cross-domain association analysis tool based on a large model, semantic association analysis is performed on the cross-domain user behavior data to generate cross-domain association analysis data. Based on the cross-domain correlation analysis data, corresponding data on user preference changes are generated.

19. The apparatus of claim 13, wherein the preference change argument countermeasure module acquires user preference reference data; The second intelligent agent performs semantic analysis on the user preference reference data, the user behavior change data, and the user preference change argument data, and executes accordingly: The evidence for each argument reflected in the data on changes in user preferences is examined. Based on the arguments rejected by the evidence review ruling, corresponding user preference change defense data for this round is generated; and / or, based on the arguments accepted by the evidence review ruling, corresponding user preference change consensus data for this round is generated.

20. The apparatus of claim 19, wherein the preference change argument countermeasure module determines, in a plurality of different user preference change types, the type of user preference change corresponding to the argument accepted by the evidence review ruling, and performs accordingly: If the corresponding user preference change type is a gradual drift type, then the consensus data of the current round of user preference change is generated accordingly to indicate that the old preferences are retained and the new preferences under observation are added. or, If the corresponding user preference change type is an event-driven mutation type, then the consensus data of the current round of user preference change is generated accordingly to indicate that a new preference has been added as the current preference; or, If the corresponding user preference change type is a scenario switching type, then the consensus data of the current round of user preference change is generated accordingly to indicate the retention of old preferences, the addition of new preferences, and the marking of applicable scenarios.

21. The apparatus of claim 19, wherein the user preference reference data includes a user preference snapshot, the user preference snapshot being a natural language description of the user preferences at a specified time. The user preference change determination module updates the user preference snapshot accordingly based on the changes in user preferences.

22. The apparatus of claim 13 or 21 further enables the first intelligent agent not to use historical user preference snapshots during the user preference analysis and processing by means of access permission control or business logic control.

23. The apparatus according to any one of claims 19 to 21, wherein the user preference reference data includes one or more user preferences and a status label corresponding to each user preference, so as to update the status accordingly based on whether the user preference has changed; At least one of the aforementioned state labels represents one of the following states: Stable, under observation, changing, controversial, historical.

24. The apparatus of claim 13, wherein the user preference change determination module determines, based on a set round threshold or time threshold, whether the processing of the first agent and the second agent in multiple rounds of the current period meets the over-threshold convergence condition; If the conditions are met, then in the current cycle, the next round of processing will not be carried out to wait for the next cycle, and it will be determined whether the user preference change argument data generated by the first agent contains any disputed arguments; If it exists, then according to the inclination of the second agent, the arguments in the dispute are temporarily processed and reserved; In the next cycle, the device, through the first and second agents, continues to process the arguments in dispute based on new user behavior change data.

25. A user preference analysis and processing device, comprising: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform: Obtain data on changes in user behavior; The first agent is determined to be configured to be inclined to believe that user preferences have changed. The first intelligent agent performs semantic analysis on the user behavior change data to generate user preference change argument data for this round. The second agent is configured to tend to deny changes in user preferences. The second intelligent agent performs semantic analysis on the user preference change argument data to generate user preference change defense data and / or user preference change consensus data in the current round. The user preference change defense data is used by the first intelligent agent to abandon the argument or generate supplementary user preference change argument data in the next round, and by the second intelligent agent to continue to defend or reach a consensus in the next round. Based on one or more rounds of processing by the first agent and the second agent, the changes in user preferences are determined.

Images