A closed-loop update method based on multi-type user behavior collection and dynamic profiling.

Abstract

This invention discloses a closed-loop update method based on multi-type user behavior collection and dynamic profile linkage, comprising: processing raw behavior data according to predefined rules to obtain the current profile update trigger event; inputting long-term historical user profile data, short-term event sequences, the current round semantic representation of the fusion context, and prompt constraints into a large language model, which outputs profile update proposal data, including label weight change values ​​and confidence levels; receiving the label weight change values ​​and confidence levels, updating and optimizing the label weight change values ​​and confidence levels through a dual constraint mechanism, storing all change records, and simultaneously generating audit logs and pushing them to the client. The dual constraint mechanism includes magnitude constraints and admission constraints. This method addresses the significant defects of existing user profile update mechanisms, where update results exhibit both jump and rigidity issues.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a closed-loop update method based on the dynamic linkage of multi-type user behavior collection and profile. Background Technology

[0002] Existing technologies, such as collaborative filtering recommendation based on user historical behavior and conventional user profiling construction solutions, are widely implemented on mainstream platforms like news reading, short videos, and content distribution. This technology revolves around data collection, offline statistics, profile generation, and business applications, with the specific process as follows:

[0003] First, the system collects basic user interaction data by setting up data points on both the client and server sides. This data includes content clicks, page exposure, page dwell time, content completion / viewing, likes, and collections. The system schedules offline ETL (Extract, Transform, Load) tasks daily to clean and aggregate the behavioral data from the past 7 days. Based on manually preset statistical rules, the system calculates metrics such as click frequency and average dwell time for each content tag to determine user interest tendencies and intensity.

[0004] After completing the statistical analysis, the system generates a flat user profile in tag-weight format and stores it in a key-value database. Profile updates typically employ either a full replacement or a sliding time window weighted approach. Finally, the recommendation engine retrieves the generated user profiles, matches them with corresponding content to complete recall and ranking, and enables personalized push notifications.

[0005] The aforementioned solutions employ two approaches: scheduled offline updates and weighted updates using sliding time windows. However, both fail to simultaneously ensure update stability and scheduling flexibility. On one hand, if users browse entirely new categories of content within a short period, the weights of user profile features will fluctuate drastically, leading to significant jumps in recommended content and severely degrading the user experience. On the other hand, long-term accumulated high-weight historical tags solidify user profiles. When user interests undergo continuous changes, the system's perception lags, its response is slow, and the output of recommended content lacks variation, resulting in poor matching. In short, the aforementioned solutions suffer from a significant flaw in their user profile update mechanism, exhibiting both abrupt changes and rigidity in the update results.

[0006] Therefore, the existing user profile update mechanism has obvious defects, and the update results have both abrupt and rigid problems, which are technical problems that urgently need to be solved in this field. Summary of the Invention

[0007] In view of this, the present invention provides a closed-loop update method based on multi-type user behavior collection and dynamic linkage of profiles, in order to solve the problem that the existing user profile update mechanism has obvious defects and the update results have both jump and rigidity problems.

[0008] This embodiment provides a closed-loop update method based on multi-type user behavior collection and dynamic profile linkage, including the following steps:

[0009] The system receives raw behavioral data reported by the client, which includes at least: reading progress data, reading duration data, dialogue content data, and dialogue sentiment data. It processes the reading progress data, reading duration data, dialogue content data, and dialogue sentiment data according to predefined rules to obtain a current profile update trigger event. The predefined rules include deep reading determination rules, quick exit determination rules, exposure without click determination rules, and dialogue preference determination rules. The current profile update trigger event includes: a deep reading event, a quick exit event, an exposure without click event, and a dialogue preference event.

[0010] The deep reading determination rule is defined as follows: if the reading progress data reaches a first preset threshold and the reading duration data reaches a second preset threshold, then the deep reading event is generated; the quick exit determination rule is defined as follows: if the reading duration data does not reach a third preset threshold and the reading progress data does not reach a fourth preset threshold, then the quick exit event is generated, wherein the first preset threshold is greater than the fourth preset threshold; the exposure without click determination rule is defined as follows: if the target content is exposed to the visible area of ​​the user terminal's display interface and there is no corresponding click behavior within a fifth preset threshold, then the exposure without click event is generated, wherein the fifth preset threshold is greater than the second preset threshold; the dialogue preference determination rule is defined as follows: parsing multi-turn dialogue semantics, identifying user preferences, and generating the dialogue preference event;

[0011] The user's long-term historical profile data, short-term event sequence, current round semantic representation of the fused context, and prompt constraints are input into the large language model. The large language model outputs profile update proposal data, which includes at least the label weight change value and confidence level. The short-term event sequence is an event sequence composed of deep reading events, quick exit events, exposure without click events, and dialogue preference events within a preset time period.

[0012] The system receives the label weight change value and the confidence level. These values ​​are updated and optimized using a dual-constraint mechanism, and a full record of all changes is stored. Simultaneously, an audit log is generated and pushed to the client. The dual-constraint mechanism includes magnitude constraints and admission constraints.

[0013] The amplitude constraint is achieved by limiting the change value of the label weight through a slow iteration strategy or a momentum-weighted update strategy to obtain the updated image label weight.

[0014] The slow iteration strategy is expressed as:

[0015] ,

[0016] In the formula, This indicates the updated image tag weight. This indicates the original image tag weight before the update. Δmax represents the label weight change value output by the large language model; Δmax represents the maximum magnitude threshold for a single update.

[0017] The momentum-weighted update strategy is expressed as follows:

[0018] ,

[0019] In the formula, This represents the change in label weights output by the large language model. Indicates the dynamic learning rate. This indicates the updated image tag weight;

[0020] The admission constraint is as follows: a confidence level execution threshold is set. Only when the confidence level is greater than or equal to the confidence level execution threshold is the image updated according to the updated image tag weight and written to the main image library; otherwise, it is only written to the main image library.

[0021] Optionally, the step of inputting long-term historical user profile data, short-term event sequences, the current round semantic representation of the fusion context, and prompt constraints into the large language model, wherein the large language model outputs profile update proposal data, including:

[0022] Based on the judgment criteria, the user's long-term historical profile data with a weight value higher than the sixth preset threshold, the profile tags corresponding to the short-term event sequence, and the topics corresponding to the current round semantic representation of the fused context are compared to obtain a set of conflict tags and a set of newly added interest tags.

[0023] Semantic integration is performed on the conflict tag set and the newly added interest tag set, and positive and negative behaviors are distinguished. The instantaneous interest intensity is obtained based on the positive and negative behaviors.

[0024] Based on the conflicting tag set, the newly added interest tag set, and the instantaneous interest intensity, the tag weight change value and the confidence level are generated.

[0025] Optionally, the determination criteria include:

[0026] If any profile tag in the user's long-term historical profile data is marked as a negative behavior more than twice in the short-term event sequence, it is determined to be a set of conflicting tags. The negative behavior includes at least the quick exit event and the exposure without click event.

[0027] If a new profile tag appears in the short-term event sequence and the corresponding weight in the user's long-term historical profile data is lower than the eighth preset threshold, it is determined to be the new interest tag set.

[0028] Optionally, the instantaneous interest intensity is expressed as:

[0029] ,

[0030] In the formula, Indicates the instantaneous intensity of interest. Indicates from =1 to Summation, Indicates will Limited to Within the interval, This represents the set of influencing factors for all relevant events in a short-term event sequence.

[0031] Optionally, the first preset threshold is 90%; the second preset threshold is 10 seconds; the third preset threshold is 5 seconds; the fourth preset threshold is 10%; and the fifth preset threshold is 5 minutes.

[0032] Optionally, the prompt constraints include: the output format is fixed structured data; the absolute value of the label weight change does not exceed 0.3; the reason for the portrait change is selected in Chinese and does not exceed 50 characters, with a confidence level of 0.0~1.0.

[0033] Compared with existing technologies, the closed-loop update method based on multi-type user behavior collection and dynamic profile linkage provided by this invention achieves at least the following beneficial effects:

[0034] First, by using a slow iteration strategy with small steps and limited amplitude or a momentum-weighted update strategy, we can avoid drastic fluctuations in user profiles caused by single, accidental raw behavioral data and prevent large jumps in recommended content. At the same time, by using a large language model to identify interest migration trends in long-term user profiles in real time, we can improve the rigidity and inflexibility of profiles, thereby balancing system stability and responsiveness to interest changes and effectively improving the user experience.

[0035] Secondly, based on predefined rules, various types of user behavior data, including positive, implicit, and semantic dimensions such as in-depth reading, quick exits, unclicked exposures, and multi-turn dialogues, are uniformly converted into current profile update trigger events, effectively overcoming the limitations of traditional solutions that can only collect single-dimensional behavior data.

[0036] Third, by mandating the storage of audit logs containing core elements such as updated profile tag weights, reasons for changes, confidence levels, and triggering events, a complete chain of evidence is established from the original behavioral data (such as clicks, browsing, and transactions) to the simultaneous writing to the main profile database or only writing to the main profile database during profile updates. This ensures that every profile evolution can be traced back to its source and deduced forward, meeting the requirements of compliance review.

[0037] Of course, any product implementing the present invention does not need to achieve all of the above additional technical effects while solving the background technical problem. It is sufficient for the product to solve the background technical problem first. The additional technical effects are effects that are beyond the understanding of those skilled in the art when the specific structure of the present invention is combined with a specific environment.

[0038] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0040] Figure 1 This is a schematic diagram of the closed-loop update method based on multi-type user behavior collection and dynamic profile linkage provided by the present invention;

[0041] Figure 2 This is a logical schematic diagram of the large language model provided by the present invention;

[0042] Figure 3 This is a comparison diagram of the innovative effects of existing technologies and the present invention. Detailed Implementation

[0043] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0044] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0045] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0046] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0047] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0048] See Figures 1 to 3 As shown, Figure 1 This is a schematic diagram of the closed-loop update method based on multi-type user behavior collection and dynamic profile linkage provided by the present invention; Figure 2 This is a logical schematic diagram of the large language model provided by the present invention; Figure 3 These are comparison images showing the innovative effects of existing technologies and the present invention in image rendering. Figure 3 In this diagram, 1 represents a jump-style user profile update, 2 represents a rigid user profile update, and 3 represents a smooth, step-by-step iterative update. This embodiment provides a closed-loop update method based on multi-type user behavior collection and dynamic user profile linkage, including the following steps:

[0049] Step S1: Receive raw behavioral data reported by the client; the raw behavioral data includes at least: reading progress data, reading duration data, dialogue content data, and dialogue sentiment data; process the reading progress data, reading duration data, dialogue content data, and dialogue sentiment data according to predefined rules to obtain the current profile update trigger event; the predefined rules include deep reading judgment rules, quick exit judgment rules, exposure without click judgment rules, and dialogue preference judgment rules; the current profile update trigger event includes: deep reading event, quick exit event, exposure without click event, and dialogue preference event; among which,

[0050] The deep reading judgment rule is defined as follows: if the reading progress data reaches the first preset threshold and the reading time data reaches the second preset threshold, a deep reading event is generated; if the reading time data does not reach the third preset threshold and the reading progress data does not reach the fourth preset threshold, a quick exit event is generated, wherein the first preset threshold is greater than the fourth preset threshold; the exposure without click judgment rule is defined as follows: if the target content is exposed to the visible area of ​​the user terminal's display interface and there is no corresponding click behavior within the fifth preset threshold, an exposure without click event is generated, wherein the fifth preset threshold is greater than the second preset threshold;

[0051] Specifically, in step S1, the server receives raw behavioral data reported by the client in real time.

[0052] When a user engages in any interactive activity on the client, such as reading, clicking, chatting, or swiping, the client collects all relevant fields in real time. In addition to regular interaction data, it simultaneously captures content attribute data, reading status data (such as reading progress data, reading duration data, reading completion rate, and bounce depth data), and chat content data. For example, all relevant fields include the user's unique identifier, operation time data, article title data, content tag data, article length data, content style data, number of images data, reading duration data, reading progress data, chat content data, and chat sentiment data.

[0053] Subsequently, the following standardized operations are performed according to the predefined data dictionary:

[0054] Field mapping: Mapping raw field names to standard internal field names (e.g., article_title) title);

[0055] Type conversion: Convert reading progress data to single-precision floating-point type float, with a valid value range of 0.0~1.0; convert reading time data to integer type int, with the unit of measurement being seconds; convert image count data to integer type int.

[0056] Unit normalization: Timestamps are uniformly converted to Unix millisecond-level integers;

[0057] Unit normalization: All time fields within the system uniformly use Unix timestamps with millisecond precision;

[0058] Enumeration value standardization: Various content styles correspond to fixed enumeration values ​​(such as rational analysis, other messages, technical analysis), and the dialogue sentiment tendency field is uniformly mapped to three fixed labels: positive (positive behavior) and negative (negative behavior).

[0059] Missing value filling: Fill missing but required fields with default values ​​(such as empty strings or 0);

[0060] Format encapsulation: The standardized fields are assembled into a structured standard JSON format, and each behavioral data is generated into an independent log line.

[0061] This ultimately generates standardized behavior logs. For example:

[0062]

[0063] It should be noted that this standardized behavior log is the raw behavior data received by the server in real time.

[0064] The server filters and transforms each piece of reading progress data, reading duration data, dialogue content data, and dialogue sentiment data according to the following steps: Specifically, the server reads the original behavior logs (such as reading progress data, reading duration data, dialogue content data, and dialogue sentiment data) one by one and checks whether they match the predefined rules.

[0065] Filter out invalid behavioral data, such as synchronously removing data with missing fields, abnormal timestamps, and conflicting behavioral logic (e.g., reading progress data is greater than 0 but page dwell time is 0), to reduce the data processing load of the subsequent large language model.

[0066] The predefined rules include rules for determining deep reading, quick exit, exposure without click, and dialogue preference. If any one of these rules is met, the data is considered valid behavioral data. If none of these rules is met, the original behavioral log is discarded and the user profile is not updated.

[0067] The deep reading determination rule is defined as follows: when the reading progress data reaches a first preset threshold and the reading time data reaches a second preset threshold, a deep reading event is generated; the first preset threshold can be 90% and the second preset threshold can be 10 seconds.

[0068] The quick exit judgment rule is defined as follows: if the reading time data does not reach the third preset threshold and the reading progress data does not reach the fourth preset threshold, a quick exit event is generated; the third preset threshold can be 5 seconds and the fourth preset threshold can be 10%.

[0069] The rule for determining if content is exposed but not clicked is defined as follows: if no corresponding click behavior occurs within a fifth preset threshold after the target content is exposed to the visible area of ​​the user terminal's display interface, an exposure-not-clicked event is generated; the fifth preset threshold can be 5 minutes.

[0070] The rule for determining if content was not clicked is implemented by writing user content exposure records to a Redis cache.

[0071] Cache key naming rules: For example, unique identifier for the user, unique identifier for the content;

[0072] The cached value stores the timestamp of this exposure.

[0073] Cache expiration time: The cache expires automatically after 300 seconds, and the exposure record is automatically cleared after expiration; if a click log is received, the cache key is deleted; if the cache key expires naturally, an exposure unclicked event is triggered.

[0074] The dialogue preference determination rule is defined as follows: parsing the semantics of multi-turn dialogues, identifying user preferences and generating dialogue preference events. The aforementioned user preferences can be the user's explicit likes or dislikes.

[0075] The specific handling method for this dialogue preference event:

[0076] Use lightweight pre-trained models (such as DistilBERT fine-tuned intent classifiers) or rule-based keyword and sentiment dictionaries;

[0077] Input the current dialogue round and the context data of the last 3 rounds, and output the topic (such as the collaboration between Retrieval Enhancement Generation (RAG) and the Agent) and the sentiment polarity (such as positive for liking and negative for rejecting).

[0078] A conversational preference event is generated only when the confidence level of the sentiment polarity is greater than 0.7 and contains explicit preference keywords (such as "like" or "don't recommend"). The event details include the topic, sentiment polarity, and original query.

[0079] It should be noted that the DistilBERT mentioned above is a lightweight distillation pre-trained model of BERT proposed in 2019. Downstream fine-tuning for intent classification is a common standard baseline solution in the fields of dialogue systems and intelligent customer service natural language processing, and belongs to existing technology.

[0080] The original paper on the DistilBERT base model is published in: Victor Sanh, Lysandre Debut, Julien Chaumond, Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter[J]. arXiv preprint, arXiv:1910.01108, October 2, 2019.

[0081] Transform into a current profile update trigger event: For each valid behavioral data point, extract key fields based on deep reading judgment rules, quick exit judgment rules, exposure without click judgment rules, and dialogue preference judgment rules, and output them according to a fixed event format. The fixed event format includes:

[0082] Unique event identifier: Globally unique string UUID, with no auto-increment conflicts; database auto-incrementing numeric primary key;

[0083] Event types: corresponding to in-depth reading judgment rules, quick exit judgment rules, exposure without click judgment rules, and dialogue preference judgment rules;

[0084] Unique user identifier, event timestamp;

[0085] Event details: Includes core data that triggers the current profile update, such as in-depth reading events carrying user topic, tags, reading progress data, and reading duration data; and exposure without clicks events carrying tags and exposure counts, etc.

[0086] Rule Identifier: Records the matching rule identifier, which is the unique number of the predefined rules that this event matched, including the deep reading judgment rule, the quick exit judgment rule, the exposure without click judgment rule, and the dialogue preference judgment rule.

[0087] It should be noted that for the rule of determining if an exposure is not clicked, an exposure cache is used with the user's unique identifier and the content's unique identifier as keys. The exposure timestamp is recorded. When click behavior data is received, the corresponding cache is cleared. The cache that has not been cleared after a timeout (e.g., more than 5 minutes) is periodically scanned, and an exposure not clicked event is generated for that exposure. Deep reading events, quick exit events, and dialogue preference events do not involve cross-log aggregation; a single valid behavior data point is used for direct determination.

[0088] The deep reading event, quick exit event, exposure unclicked event, and dialogue preference event output in step S1 will be directly input into step S2 to reduce invalid calls.

[0089] Step S2: Input the user's long-term historical profile data, short-term event sequence, current round semantic representation of the fusion context, and prompt constraints into the large language model. The large language model outputs profile update proposal data, which includes at least the label weight change value and confidence level. The short-term event sequence is an event sequence composed of deep reading events, quick exit events, exposure without click events, and dialogue preference events within a preset time period.

[0090] Specifically, the user's long-term historical profile data in step S2 is the latest version of tag weight data stored in the main profile database.

[0091] The latest version tag weight data refers to the set of tag weights that were most recently successfully updated and written to the main profile database. Each tag (such as "AI (Artificial Intelligence)") corresponds to a weight value (e.g., 0.7). This latest version tag weight data is generated by the slow iteration strategy or momentum-weighted update strategy in step S3, and is a globally unique latest version.

[0092] In step S2, the preset time period is within 24 hours, and the short-term event sequence consists of deep reading events, quick exit events, exposure without click events, and dialogue preference events generated by the user in the last 24 hours.

[0093] The short-term event sequence in step S2 specifically includes the current profile update trigger event output in step S1. The event types cover deep reading events, quick exit events, exposure without click events, and dialogue preference events. Each deep reading event, quick exit event, exposure without click event, and dialogue preference event includes an event type, a unique user identifier, an event occurrence timestamp, event details (such as deep reading events carrying user topic, tags, reading progress data, and reading duration data), and a rule identifier.

[0094] In step S2, the semantic representation of the current round of the fused context is: the semantic parsing results of the most recent rounds of dialogue in the current dialogue session.

[0095] Optionally, see Figure 2 As shown, the user's long-term historical profile data, short-term event sequences, the current round semantic representation of the fused context, and the prompt constraints are input into the large language model. The large language model outputs profile update proposal data, including:

[0096] Step S21: Based on the judgment conditions, the user's long-term historical profile data with a weight value higher than the sixth preset threshold, the profile tags corresponding to short-term event sequences, and the topics corresponding to the current round semantic representation of the fusion context are compared to obtain a set of conflicting tags and a set of newly added interest tags.

[0097] It should be noted that the constraints in step S21 include: the data must not change with the input data; the output format must be fixed structured data (such as a fixed JSON structure); the absolute value of the label weight change must not exceed 0.3; the reason for the profile change must be in Chinese and not exceed 50 characters, with a confidence level of 0.0~1.0. These constraints are used to ensure the accuracy of the profile update proposal data as much as possible and to avoid arbitrary generation.

[0098] In step S21, the sixth preset threshold can be 0.5. The user's long-term historical profile data with a weight value higher than 0.5, the profile tags appearing in the short-term event sequence, and the topics in the current round of semantic representation of the fusion context are compared together.

[0099] Optionally, see Figure 2 As shown, the judgment criteria include: if any profile tag in the user's long-term historical profile data is marked as a quick bounce event or an exposure without click event more than twice in a short-term current profile update trigger event, it is judged as a conflict tag set;

[0100] If a new profile tag appears in a short-term event sequence and the corresponding weight in the user's long-term historical profile data is lower than the eighth preset threshold, it is determined to be a new set of interest tags. The eighth preset threshold can be 0.1.

[0101] Judgment criteria: If any profile tag in the user's long-term historical profile data is marked as a negative behavior more than twice in a short-term event sequence (such as a quick bounce event or an exposure without click event), it is judged as a conflicting tag set; if a new profile tag appears in a short-term event sequence and the corresponding weight value of the profile tag in the user's long-term historical profile data is less than 0.1, it is judged as a "new interest tag set"; if neither of the above conditions applies, it is judged as "consistent".

[0102] Using the above scheme, the conflict tag set and the newly added interest tag set are determined according to the judgment conditions, providing the prerequisites for step S21.

[0103] Step S22: Semantically integrate the conflict tag set and the newly added interest tag set, distinguish between positive and negative behaviors, and obtain the instantaneous interest intensity based on the positive and negative behaviors;

[0104] Association: The conflict tag set and the newly added interest tag set output in step S21 are used as the input in step S22. Only the conflict tag set and the newly added interest tag set are semantically integrated. There is no need to process the profile tags that have been determined to be consistent.

[0105] Semantic integration of conflict tag sets and newly added interest tag sets specifically includes: collecting all events related to the target tag in short-term event sequences, and mapping the event type of each event to an impact factor: for example, deep reading events are mapped to +0.2; quick exit events are mapped to -0.15; exposure without clicks events are mapped to -0.05; dialogue preference events are mapped to 0.25 (positive behavior); and dialogue preference events are mapped to -0.3 (negative behavior).

[0106] Distinguish between positive and negative behaviors: Positive behaviors refer to deep reading events and conversation preference events. Negative behaviors refer to quick exit events, unclicked exposure events, and conversation preference events.

[0107] Calculate the instantaneous interest intensity of each profile tag: For each profile tag, sum the influence factors of all its related events and limit them to the interval [-0.5, +0.5] to obtain the instantaneous interest intensity, i.e., the original calculated value of LLMΔ.

[0108] Step S23: Based on the conflicting tag set, the newly added interest tag set, and the instantaneous interest intensity, generate the tag weight change value, the reason for the change, and the confidence level.

[0109] The example corresponding to step S23 is as follows:

[0110]

[0111] Through steps S21 to S23, accurate identification of user intent is achieved.

[0112] Step S3: Receive the label weight change value and confidence level. The label weight change value and confidence level are updated and optimized through a dual constraint mechanism, and all change records are stored. Simultaneously, an audit log is generated and pushed to the client. The dual constraint mechanism includes magnitude constraint and admission constraint.

[0113] Amplitude constraint: Limit the amplitude of the label weight change value through a slow iteration strategy or a momentum-weighted update strategy to obtain the updated image label weight;

[0114] The slow iteration strategy is represented as:

[0115] ,

[0116] In the formula, This indicates the updated image tag weight. This indicates the original image tag weight before the update. This represents the label weight change value output by the large language model; Δmax represents the maximum magnitude threshold for a single update.

[0117] The momentum-weighted update strategy is expressed as:

[0118] ,

[0119] In the formula, This represents the change in label weights output by the large language model. Indicates the dynamic learning rate. This indicates the updated image tag weight;

[0120] Admission constraints: Set a confidence threshold. Only when the confidence level is greater than or equal to the confidence threshold will the image be updated according to the updated image tag weight and written to the main image library; otherwise, it will only be written to the main image library.

[0121] Specifically, the system does not directly execute the profile update proposal data output in step S2, but optimizes the update result through a dual constraint mechanism while storing the full change record.

[0122] The label weight change values ​​output by the above large language model are truncated by the maximum single update amplitude threshold Δmax, and then the weights are updated by addition. This is a constrained online learning update algorithm that ensures that the single-step update amplitude is controllable.

[0123] The dual constraint mechanism in step S3 includes amplitude constraint and admission constraint. The confidence threshold in the admission constraint can be 0.6. The amplitude constraint and admission constraint together constitute the optimization constraint on the output of the large language model.

[0124] Slow iteration strategy or momentum-weighted update strategy applies to the change value of tag weight, and admission constraint applies to whether to perform profile update.

[0125] Application Example 1: Positive Behavior Interest Update:

[0126] Basic parameters: In the slow iteration strategy, the original profile label weights before the user profile is updated are: Technology = 0.7, Entertainment = 0.2; the maximum update amplitude per cycle is Δmax = 0.15, and the confidence threshold is T = 0.6.

[0127] The raw behavioral data in step S1: Read a history and culture article completely, with a reading progress of 98% and a reading time of 300 seconds, generating a deep reading event.

[0128] In step S2, the large language model outputs profile update proposal data: historical tag weight change value LLMΔ=+0.2, confidence level=0.9, reason for change: users have read historical content in depth and show obvious interest.

[0129] Step S3 Formula calculation: ,therefore, Confidence level: 0.9 > 0.6, execute profile update.

[0130] End-user profile: Technology = 0.7, Entertainment = 0.2, History = 0.15. All change records are synchronously written, and subsequent audit logs fully record the triggering events, reasons for changes, and confidence levels.

[0131] Implementation results: The main profile is updated positively based on the deep reading event. The update range is stable without sudden changes. The whole process is explainable and traceable, and the basic functions are running normally.

[0132] Application Example 2: Negative Interest Decay;

[0133] Basic parameters: In the slow iteration strategy, the original profile label weights before the user profile is updated are: AI=0.8, Technology=0.6; Δmax=0.15, confidence threshold T=0.6.

[0134] The raw behavioral data in step S1: three views of AI-related content were not clicked, generating an exposure-without-click event. In addition, the user explicitly stated in the dialogue preferences that they would no longer receive AI-related content, generating a dialogue preference event. This dialogue preference event corresponds to negative behavior.

[0135] Step S2 outputs profile update proposal data from the large language model: AI label weight change values. Confidence level = 0.88, Reason for change: Users repeatedly ignored AI content, and combined with the dialogue, their interest in this type of content has significantly decreased.

[0136] Step S3 Formula calculation: ,therefore, With a confidence level of 0.88 > 0.6, perform a profile update.

[0137] Final user profile: AI = 0.65, Technology = 0.6. A full change log is simultaneously written, and subsequent audit logs fully record the triggering events, reasons for changes, and confidence levels.

[0138] Implementation Results: By fully leveraging behaviors that traditional technologies cannot identify, such as unclicked exposures and multi-turn conversations, and appropriately reducing the weight of outdated interests, the problem of rigid user profiles was solved. The update process was smooth, and all operations are traceable.

[0139] Application Example 3: A specific example of the momentum-weighted update strategy:

[0140] Assume the user's current "AI" tag has an original weight of Wold=0.70;

[0141] The label weight change value output by the large language model: Wpro=0.90 (indicating suggested enhancement); learning rate α=0.15;

[0142] Substitution expression:

[0143] .

[0144] Results: The portrait weight was updated from 0.70 to 0.73, which is smoother than setting it directly to 0.90, preserves historical information, and avoids sudden changes in the portrait caused by fluctuations in the output of a single large language model.

[0145] Whether the updated profile tag weight has changed in step S3, the system will uniformly record: operation time data, user unique identifier, changed field, original value, updated profile tag weight, change range, change reason, confidence level, and the unique identifier of the event that triggered this update, forming a complete and traceable data link.

[0146] See Figure 3 As shown, the horizontal axis represents time, and the vertical axis represents the image tag weight. The update curves of existing technologies exhibit a jump pattern of sharp increases and decreases (e.g., 1 corresponds to the red line), or a rigid pattern that remains unchanged for a long time (e.g., 2 corresponds to the yellow line); the curve in this embodiment is a smooth step change, and the maximum amplitude threshold of a single update is controllable, which can avoid data mutations and respond promptly to changes in users' real interests.

[0147] It should be noted that the large language model in this example can be the Doubao Seed 2.0 engine, or the DeepSeek model. This example does not make any specific restrictions on this.

[0148] Compared with existing technologies, the closed-loop update method based on multi-type user behavior collection and dynamic profile linkage provided in this embodiment achieves at least the following beneficial effects:

[0149] This embodiment provides a closed-loop update method based on the dynamic linkage of multi-type user behavior collection and profile.

[0150] First, by using a slow iteration strategy with small steps and limited amplitude or a momentum-weighted update strategy, we can avoid drastic fluctuations in user profiles caused by single, accidental raw behavioral data and prevent large jumps in recommended content. At the same time, by using a large language model to identify interest migration trends in long-term user profiles in real time, we can improve the rigidity and inflexibility of profiles, thereby balancing system stability and responsiveness to interest changes and effectively improving the user experience.

[0151] Secondly, based on predefined rules, various types of user behavior data, including positive, implicit, and semantic dimensions such as in-depth reading, quick exits, unclicked exposures, and multi-turn dialogues, are uniformly converted into current profile update trigger events, effectively overcoming the limitations of traditional solutions that can only collect single-dimensional behavior data.

[0152] Third, by mandating the storage of audit logs containing core elements such as updated profile tag weights, reasons for changes, confidence levels, and triggering events, a complete chain of evidence is established from the original behavioral data (such as clicks, browsing, and transactions) to the simultaneous writing to the main profile database or only writing to the main profile database during profile updates. This ensures that every profile evolution can be traced back to its source and deduced forward, meeting the requirements of compliance review.

[0153] It's important to note that the aforementioned implicit behaviors refer to behavioral traces embedded in a user's interaction trajectory, rather than explicitly expressed through actions like liking, rating, saving, or commenting. Typical examples include reading progress data, page dwell time, unclicked exposures, rapid exits, and emotional inclinations in conversations (such as the negative attitude implied by "interest waning"). These signals require contextual or semantic analysis for effective identification and quantification, serving as an important supplement to explicit feedback.

[0154] In one alternative embodiment, the instantaneous interest intensity is represented as:

[0155] ,

[0156] In the formula, Indicates the instantaneous intensity of interest. Indicates from =1 to Summation, Indicates will Limited to Within the interval, This represents the set of influencing factors for all relevant events in a short-term event sequence.

[0157] Specifically, the set of influence factors for all relevant events in a short-term event sequence for any given profile tag is: ,but:

[0158] ,

[0159] In the formula, Indicates: If Then take ;like Then take Otherwise take .

[0160] Application example 1 of instantaneous interest intensity:

[0161] Historical profile update trigger events related to the profile tag "AI": 1 deep reading event (+0.2), 1 dialogue preference event (positive behavior) (+0.25). =+0.45, in Within the range, the upper limit was not exceeded, p=0.45.

[0162] Verification: The expression corresponding to the instantaneous interest intensity is reasonable, and the output is in Inside.

[0163] Application example 2 of instantaneous interest intensity:

[0164] Historical portrait updates related to the "Entertainment" tag trigger events: 3 quick pop-up events One exposure without clicks (-0.05). =-0.5, not exceeding the lower limit; add 1 more dialogue preference event (negative behavior) -0.3, =-0.8, triggering the lower limit limiting, resulting in .

[0165] Verification: The amplitude limiting mechanism is effective, preventing excessively large single update amplitudes, which meets the requirements of the slow iteration strategy.

[0166] Using the above scheme, based on the set of influence factors of all relevant events in a short-term event sequence for any profile tag, and... The instantaneous interest intensity is obtained and used as the label weight change value output by the large language model.

[0167] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0168] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A closed-loop update method based on multi-type user behavior collection and dynamic profile linkage, characterized in that, Includes the following steps: The system receives raw behavioral data reported by the client, which includes at least: reading progress data, reading duration data, dialogue content data, and dialogue sentiment data. It processes the reading progress data, reading duration data, dialogue content data, and dialogue sentiment data according to predefined rules to obtain a current profile update trigger event. The predefined rules include deep reading determination rules, quick exit determination rules, exposure without click determination rules, and dialogue preference determination rules. The current profile update trigger event includes: a deep reading event, a quick exit event, an exposure without click event, and a dialogue preference event. The deep reading determination rule is defined as follows: if the reading progress data reaches a first preset threshold and the reading duration data reaches a second preset threshold, then the deep reading event is generated; the quick exit determination rule is defined as follows: if the reading duration data does not reach a third preset threshold and the reading progress data does not reach a fourth preset threshold, then the quick exit event is generated, wherein the first preset threshold is greater than the fourth preset threshold; the exposure without click determination rule is defined as follows: if the target content is exposed to the visible area of ​​the user terminal's display interface and there is no corresponding click behavior within a fifth preset threshold, then the exposure without click event is generated, wherein the fifth preset threshold is greater than the second preset threshold; the dialogue preference determination rule is defined as follows: parsing multi-turn dialogue semantics, identifying user preferences, and generating the dialogue preference event; The user's long-term historical profile data, short-term event sequence, current round semantic representation of the fused context, and prompt constraints are input into the large language model. The large language model outputs profile update proposal data, which includes at least the label weight change value and confidence level. The short-term event sequence is an event sequence composed of deep reading events, quick exit events, exposure without click events, and dialogue preference events within a preset time period. The system receives the label weight change value and the confidence level. These values ​​are updated and optimized using a dual-constraint mechanism, and a full record of all changes is stored. Simultaneously, an audit log is generated and pushed to the client. The dual-constraint mechanism includes magnitude constraints and admission constraints. The amplitude constraint is achieved by limiting the change value of the label weight through a slow iteration strategy or a momentum-weighted update strategy to obtain the updated image label weight. The slow iteration strategy is expressed as: ， In the formula, This indicates the updated image tag weight. This indicates the original image tag weight before the update. Δmax represents the label weight change value output by the large language model; Δmax represents the maximum magnitude threshold for a single update. The momentum-weighted update strategy is expressed as follows: ， In the formula, This represents the change in label weights output by the large language model. Indicates the dynamic learning rate. This indicates the updated image tag weight; The admission constraint is as follows: a confidence level execution threshold is set. Only when the confidence level is greater than or equal to the confidence level execution threshold is the image updated according to the updated image tag weight and written to the main image library; otherwise, it is only written to the main image library.

2. The closed-loop update method based on multi-type user behavior collection and dynamic profile linkage according to claim 1, characterized in that, The process involves inputting long-term historical user profile data, short-term event sequences, the current round semantic representation of the fused context, and prompt constraints into a large language model. The large language model then outputs profile update proposal data, including: Based on the judgment criteria, the user's long-term historical profile data with a weight value higher than the sixth preset threshold, the profile tags corresponding to the short-term event sequence, and the topics corresponding to the current round semantic representation of the fused context are compared to obtain a set of conflict tags and a set of newly added interest tags. Semantic integration is performed on the conflict tag set and the newly added interest tag set, and positive and negative behaviors are distinguished. The instantaneous interest intensity is obtained based on the positive and negative behaviors. Based on the conflicting tag set, the newly added interest tag set, and the instantaneous interest intensity, the tag weight change value and the confidence level are generated.

3. The closed-loop update method based on multi-type user behavior collection and dynamic profile linkage according to claim 2, characterized in that, The determination criteria include: If any profile tag in the user's long-term historical profile data is marked as a negative behavior more than twice in the short-term event sequence, it is determined to be a set of conflicting tags. The negative behavior includes at least the quick exit event and the exposure without click event. If a new profile tag appears in the short-term event sequence and the corresponding weight in the user's long-term historical profile data is lower than the eighth preset threshold, it is determined to be the new interest tag set.

4. The closed-loop update method based on multi-type user behavior collection and dynamic profile linkage according to claim 2, characterized in that, The instantaneous interest intensity is expressed as: ， In the formula, Indicates the instantaneous intensity of interest. Indicates from =1 to Summation, Indicates will Limited to Within the interval, This represents the set of influencing factors for all relevant events in a short-term event sequence.

5. The closed-loop update method based on multi-type user behavior collection and dynamic profile linkage according to claim 1, characterized in that, The first preset threshold is 90%; the second preset threshold is 10 seconds; the third preset threshold is 5 seconds; the fourth preset threshold is 10%; and the fifth preset threshold is 5 minutes.

6. The closed-loop update method based on multi-type user behavior collection and dynamic profile linkage according to claim 1, characterized in that, The prompt constraints include: the output format is fixed structured data; the absolute value of the label weight change does not exceed 0.3; the reason for the profile change is selected in Chinese and does not exceed 50 characters, with a confidence level of 0.0~1.0.

Images