A real-time information evolution prediction method and system based on user portrait

Abstract

The present application relates to the technical field of information prediction, in particular to a real-time information evolution prediction method and system based on user portrait, the method constructs an interest trend vector by performing semantic embedding and weighted calculation on user behavior sequence, forming a dynamic portrait structure that can reflect the user interest migration rate; the multi-version semantic sequence of information content is guided and screened by using the user interest trend, an information semantic evolution trend vector sensitive to the user is constructed, and the personalized direction alignment of information evolution modeling is realized; by designing a trend resonance scoring function, the user interest direction, interest change rate and information evolution direction are jointly modeled, the trend drift response term and semantic stability regularization term are introduced, and the accuracy and robustness of matching are improved; a response probability prediction model is established, and the results are used to drive system pushing, public opinion marking or supervision warning operations, realizing the whole process linkage from interest evolution to behavior prediction.

Description

Technical Field

[0001] This invention relates to the technical field of information prediction, and more particularly to a method and system for predicting real-time information evolution based on user profiles. Background Technology

[0002] With the accelerated speed of information dissemination and the increasing complexity of multi-source content ecosystems, the amount of information users encounter daily is exploding, requiring platforms to complete more refined content filtering and personalized push notifications in a shorter time.

[0003] Existing systems generally rely on user profiling technology, building user models by collecting users' historical behavior, preferences, and interest tags for content recommendation or public opinion analysis. However, traditional user profiles are often static or updated infrequently, failing to reflect the changing patterns of user interests over time. Meanwhile, the information content itself is constantly evolving. Multiple versions of news reports, semantic drift in social media topics, and the timely diffusion of content context mean that "information" is no longer fixed text but a dynamically evolving semantic flow. Existing methods typically model both users and information independently. User models are often built based on historical click frequency or collaborative filtering, while information models are achieved through keyword statistics or popularity regression. The two lack temporal correlation and dynamic coupling mechanisms. When information undergoes version updates or semantic shifts, the system struggles to determine whether users still maintain interest or predict whether they will generate new responses in the future. Furthermore, traditional recommendation algorithms only focus on static interest matching, ignoring the trend characteristics of interests, such as the shift in user interest from "new energy vehicles" to "hybrid energy systems," or the evolution of information content from "macroeconomics" to "high-end manufacturing." This leads to a disconnect between the recommended results and the user's actual attention direction. Especially in highly dynamic scenarios such as news recommendation, public opinion tracking, and intelligent supervision, existing methods cannot achieve real-time prediction and personalized trend inference.

[0004] In summary, existing technologies have two main shortcomings: first, they lack trend expression of user interests, making it difficult to capture their dynamic migration direction; second, they lack semantic path modeling of information evolution, making it impossible to quantify its evolutionary trend toward user interests. This disconnect makes it impossible to accurately characterize the dynamic resonance relationship between users and content, thus limiting the system's ability to predict future information response behavior. Summary of the Invention

[0005] This invention proposes a real-time information evolution prediction method and system based on user profiles, which innovatively realizes a two-way dynamic modeling and matching mechanism for user interest trends and information semantic evolution, thereby constructing a trend-aware closed-loop system that can predict users' future response behavior in real time.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: On the one hand, a method for predicting real-time information evolution based on user profiles is provided, including the following steps: The system collects and sorts user content interaction behaviors in real time through a fixed-length sliding window. The content involved in each behavior is input into the pre-trained BERT-base model, which outputs a semantic vector. The system combines a weight mechanism to construct a user interest trend vector. Finally, through a time difference comparison mechanism, the system selects a preset time point to construct the interest trend vector of the previous period and calculates the change in the current user's interest direction in the time dimension. Select all information items with semantic multiple version records from the current candidate information set, input them into the BERT-base model for vector encoding, obtain semantic representation, and use the user interest trend vector as the filtering function in the content trend extraction process. Score the directional consistency of each pair of adjacent semantic change segments. By setting a threshold, only retain all version change segments with consistency scores greater than the threshold, and sum them according to their strength to construct the semantic evolution trend vector of each candidate information that is sensitive to the current user. The trend matching function measures the similarity between a piece of information and the user's current interest direction, and evaluates whether the information trend vector is moving closer to the direction of interest change. Finally, a trend alignment score is obtained, which is used to quantify the degree of trend resonance between each candidate piece of information and the current user profile. The higher the trend alignment score, the more likely the information is evolving into content that the user is potentially interested in. By using the response probability prediction formula and considering factors such as the stability and timeliness of information evolution, a personalized response probability estimate is performed for each candidate message. Candidate messages with response probabilities greater than a preset threshold are then selected and corresponding personalized message pushes are executed.

[0007] Preferably, the pre-trained BERT-base model includes 12 Transformer layers with a hidden size of 768, used to concatenate the content title and body text, and normalizes the vectors during the encoding stage.

[0008] Preferably, the content interaction behavior includes content ID, behavior type, timestamp, page dwell time, and viewing ratio, wherein the behavior type includes browsing, clicking, favorite, and commenting.

[0009] Preferably, the directional consistency scoring includes calculating the cosine score of the angle between different versions of the semantic change vector of each piece of information content, which is used to represent the resonance intensity. In the process of calculating the cosine score, there are basic directional consistency scores and innovative dynamic trend guidance items, which are used to measure the consistency between the content evolution direction and the current interest trend and adjust the influence ratio between the current interest and the change trend, respectively.

[0010] Preferably, in the process of constructing the semantic evolution trend vector, by filtering out content change segments that do not meet the interest resonance threshold, the user's interest direction and its rate of change are introduced as a filtering method for content evolution trend extraction, and finally a semantic evolution trend vector is generated.

[0011] Preferably, the trend matching function includes a static interest alignment term, an interest drift response term, and a trend direction fluctuation penalty term. The static interest alignment term measures whether the current interest direction is consistent with the content evolution direction. The interest drift response term measures whether the content is conforming to the user's interest drift trend. If the value of the interest drift response term is higher, it indicates that the content trend has a leading effect and meets the prediction target. The trend direction fluctuation penalty term measures whether the changes in the content in the semantic space are consistent in direction, reflects the stability of the content evolution, and avoids high-frequency fluctuating content being misjudged as high trend resonance in the matching score.

[0012] Preferably, the lower the trend alignment score, the less interested the user is in the information.

[0013] Preferably, the response probability prediction formula constructs a prediction model using trend alignment score, information evolution uncertainty coefficient, and information timeliness index, and outputs the response probability of each candidate piece of information within a future time window.

[0014] Preferably, the actual operation includes pushing the information to the user's homepage or topic stream in the recommendation scenario, triggering keyword highlighting or sentiment polarity classification in the risk monitoring scenario, and attaching it to the event evolution node and initiating emergency response logic in the digital twin supervision system.

[0015] In another aspect of the present invention, a real-time information evolution prediction system based on user profiles is provided, comprising: connected in sequence: The data acquisition and processing module is used to collect and sort user content interaction behaviors in real time through a fixed-length sliding window. The content involved in each behavior is input into the pre-trained BERT-base model to output semantic vectors. Combined with the weight mechanism, a user interest trend vector is constructed. Finally, through the time difference comparison mechanism, a preset time point is selected to construct the interest trend vector of the previous period and calculate the change of the current user's interest direction in the time dimension. The information evolution trend vector construction module is used to select all information items with semantic multi-version records from the current candidate information set, input them into the BERT-base model for vector encoding, obtain semantic representation, and use the user interest trend vector as the filtering function in the content trend extraction process. It scores the directional consistency of each pair of adjacent semantic change segments, and by setting a threshold, only retains all version change segments with consistency scores greater than the threshold, and sums them according to their strength to construct the semantic evolution trend vector of each candidate information that is sensitive to the current user. The personalized information trend assessment module is used to measure the similarity between a piece of information and the user's current interest direction through a trend matching function, and to assess whether the information trend vector is moving closer to the direction of interest change, and finally obtains a trend alignment score, which is used to quantify the degree of trend resonance between each candidate piece of information and the current user profile. The higher the trend alignment score, the more likely the information is evolving into content that the user is potentially interested in. The trend matching probability prediction module is used to estimate the personalized response probability of each candidate message by combining the response probability prediction formula with the stability and timeliness of information evolution factors. It then filters candidate messages with response probabilities greater than a preset threshold and executes corresponding personalized message pushes.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a real-time information evolution prediction method and system based on user profiles. It innovatively realizes a bidirectional dynamic modeling and matching mechanism for user interest trends and information semantic evolution, thereby constructing a trend-aware closed-loop system capable of predicting future user response behaviors in real time. First, by semantically embedding and weighting user behavior sequences, the invention constructs an interest trend vector with directional and inertial expression capabilities, forming a dynamic profile structure that reflects the rate of user interest migration. Second, it uses user interest trends to guide the filtering of multi-version semantic sequences of information content, constructing a user-sensitive information semantic evolution trend vector, achieving personalized directional alignment in information evolution modeling. Third, by designing a trend resonance scoring function, it jointly models user interest direction, interest change rate, and information evolution direction, introducing trend drift response terms and semantic stability regularization terms to improve the accuracy and robustness of matching. Finally, based on trend matching scores, information uncertainty coefficients, and timeliness indicators, a response probability prediction model is established, and the results are used to drive system push, public opinion labeling, or regulatory early warning operations, realizing full-process linkage from interest evolution to behavior prediction. This invention solves the problems of the inability to express user interests in real time, the inability to quantify information evolution, and the lack of trend coupling between users and information by organically coordinating four steps: "user trend modeling, content evolution extraction, trend alignment calculation, and behavior prediction execution." It provides a technical solution with high timeliness, high interpretability, and high prediction accuracy for dynamic content distribution and intelligent supervision systems. Attached Figure Description

[0017] Figure 1 This is a flowchart of a real-time information evolution prediction method based on user profiles in a specific embodiment of the present invention; Figure 2 This is a block diagram of a real-time information evolution prediction system based on user profiles, as described in a specific embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention aims to construct interest trend vectors based on users' recent behavior on a platform, expressing the dominant direction of a user's interests within the current time period. Unlike traditional static user profiles based on tags or content categories, this vector is dynamically extracted from the user's temporal behavior, reflecting the real-time changing trends of user interests and thus supporting subsequent information evolution trend matching. This invention deals with a rapidly evolving real-time environment, therefore, user profiles must also be time-sensitive.

[0020] Please refer to Figure 1 As shown, this application proposes a real-time information evolution prediction method based on user profiles, including the following steps: Step 1: Real-time collection and sorting of user content interaction behaviors through a fixed-length sliding window. The content involved in each behavior is input into a pre-trained BERT-base model, which outputs a semantic vector. A user interest trend vector is constructed using a weighting mechanism. Finally, through a time difference comparison mechanism, a preset time point is selected to construct the interest trend vector of the previous period and calculate the change in the current user's interest direction over time. Specifically, this includes: This step uses the behavior sequence of a sliding time window as input, and combines content semantic vector extraction and behavior weighting mechanism to form a directional user interest representation.

[0021] Input data comes from a behavior collection SDK embedded in the client (such as an app, mini-program, or webpage). The system records every user content interaction in real time, including content ID, behavior type (such as browsing, clicking, favoriting, commenting), timestamp, page dwell time, and viewing ratio. The data enters the Flink stream processing platform via a Kafka message queue, where it is filtered, cleaned, and structured to form a unified format sequence of behavior records, represented as... ,in For user ID, For content ID, For timestamps, Encode the behavior type.

[0022] The platform uses a fixed-length sliding window (such as the last 15 minutes or 10 actions) to extract and sort the user's action sequence. Each action involves... The input is a pre-trained BERT-base model (12-layer Transformer, hidden size 768). After concatenating the title and body text, the [CLS] bits are extracted as the semantic vector. This model is a publicly released Chinese model, which has been fine-tuned according to standards, and the vectors have been normalized during the encoding stage.

[0023] To quantify the impact of different behaviors on user interest, a weighting mechanism is introduced. Behavior type A predefined weight table is used, for example, views = 0.3, clicks = 0.6, favorites = 0.9, comments = 1.0, and the final weight coefficient is calculated by normalization based on the dwell time corresponding to the behavior. The user interest trend vector is then constructed as follows:

[0024] in, This represents the current user's interest trend vector. The number of actions within the sliding time window. The weight coefficient for the i-th behavior (determined jointly by the behavior type and duration of stay). Let i be the semantic encoding vector of the i-th content. This is a normalization factor that makes the representations of different users in the entire vector space comparable.

[0025] In actual operation, for example, a user might read three tech news articles and save a chip-related feature article within the last 15 minutes, then briefly browse an entertainment article. Because the saving behavior and the longer viewing time for tech content result in higher engagement, the corresponding... The higher the value, the more the synthesized vectors will be biased towards the scientific and technological semantic space. This type of trend vector is better at capturing users' attention trends within a micro-timeframe than traditional interest tags.

[0026] To further indicate whether interest is shifting, the system introduces a time difference comparison mechanism based on the sliding window, selecting a specific time point. (e.g., 5 minutes ago) Construct the interest trend vector for the previous time period. And calculate the direction of change of current interest:

[0027] in, It represents the amount of change in a user's interests over time, reflecting whether there is a significant trend of interest drift. Calculated using the same method as the current time, with input being records of actions from a slightly earlier time window. Through... By determining the norm or direction, the system can classify the current state of interest (such as stable, transitional, fluctuating, etc.), providing dynamic guidance for the extraction of information trend direction in subsequent steps.

[0028] Book and All vectors are normalized and used for personalized modeling and matching calculations of subsequent information evolution trends. This modeling approach, by introducing the temporal and semantic spatial directionality of user behavior, breaks free from the limitations of static tags or content click frequency, enabling user profiles to be expressed in real-time, directional, and matchable, forming the starting point of the bidirectional evolution model of users and information in this invention.

[0029] Step 2: Select all information items with semantically multiple versions from the current candidate information set, input them into the BERT-base model for vector encoding, obtain semantic representations, and use the user interest trend vector as a filtering function in the content trend extraction process. For each pair of adjacent semantic change segments, score the directional consistency. By setting a threshold, only retain all version change segments with consistency scores greater than the threshold, and sum them according to their strength to construct a semantic evolution trend vector that is sensitive to the current user for each candidate information item. Specifically, this includes: To achieve real-time information evolution prediction based on user profiles, this step revolves around a core objective: to construct a "user-sensitive semantic evolution trend vector" that highly matches user interest trends in the context of multiple versions of semantic evolution of information content, for subsequent trend-level alignment and prediction with user interest directions.

[0030] This goal is achieved by using the two vectors output in step one: the user's current main direction of interest vector. and the rate of change of user interests .in, Used for filtering and extracting content to guide its evolution, It then participates in the adjustment of evolution intensity as a dynamic weight.

[0031] Select all information items with semantically multi-version records from the current candidate information set. Each piece of information contains at least two or more time-based versions of the text, such as initial news drafts, editorial revisions, user comment summaries, and aggregated public opinion summaries. In real-world platforms, this version data can be obtained through the version control interface of a news content management system (CMS), the behavior log reverse lookup mechanism of a content distribution platform (CDP), and calls to social media aggregation services. Typically, each piece of information can collect [number of versions]. A text version at different points in time. For each version's content... Vector encoding is performed using the same BERT-base model (Unified Semantic Space) as in step one to obtain semantic representations. The vector output is normalized.

[0032] To construct a semantic evolution trend vector, instead of using ordinary time series differencing modeling, a user interest main direction vector is introduced. As a filtering function in the content trend extraction process, it scores the directional consistency of each pair of adjacent semantic change segments, thus retaining only those parts that "evolve along the direction of user interests." Information content is derived from version... The semantic change vector to version k is represented as follows Then for each segment and Calculate the cosine score of the included angle Characterizing resonance intensity:

[0033] In this formula, the first term is the basic directional consistency score, which measures whether the direction of content evolution is consistent with the current interest trend; the second term introduces an innovative dynamic trend-oriented term, using the rate of change of user interests. To guide whether information content aligns with user attention trends; This is a balancing coefficient (set empirically, such as 0.5), used to adjust the proportion of influence between current interests and changing trends.

[0034] Furthermore, to ensure that the constructed trend vector has directional expressive power and is unaffected by noise, a threshold is set. (e.g., 0.3), only keep all Version changes, and according to their intensity Weighted summation to construct information semantic trend vector : ; in, It is a normalization constant (i.e., the sum of the weights of all retained segments). This is an indicator function used to filter content change segments that do not meet the interest resonance threshold. .

[0035] The innovation of this structure lies in the fact that traditional content evolution modeling often treats all version difference information equally, while the present invention introduces user interest direction and its rate of change as a "filter for content evolution trend extraction" for the first time, so that the final generated content trend vector has personalization, directionality and filtering, and can more accurately capture the "content that is becoming important to the user".

[0036] In practical implementation, let's take a real-world platform scenario as an example: A user's current interest leans towards "domestic high-end manufacturing," and recently, this interest has gradually shifted from general finance to "new energy equipment." Meanwhile, a news article's content evolved from reporting on the "US-China Economic Meeting" on the first page, to adding "Chinese manufacturers speaking" on the second page, and then inserting "new energy vehicle battery issues" on the third page. This illustrates the direction of the third version update for this information. With changes in user interests This will generate a high resonance value, so this segment is retained and its weight in the final trend vector is enhanced, increasing its likelihood of being responded to by users in the future.

[0037] The output of this step is each candidate message. The "semantic evolution trend vector sensitive to the current user" This vector has undergone triple personalization: guidance direction alignment, interest change enhancement, and content noise suppression. It possesses high semantic relevance and trend expression capabilities, and is one of the sole input sources for interest-information trend coupling modeling in subsequent steps. It constitutes the core of the user profile-guided information modeling innovation link in the entire invention scheme structure.

[0038] Step 3: Measure the similarity between a piece of information and the user's current interest direction using a trend matching function, and evaluate whether the information trend vector is moving closer to the direction of interest change. This results in a trend alignment score, which quantifies the degree of trend resonance between each candidate piece of information and the current user profile. A higher trend alignment score indicates that the information is evolving into content that the user is potentially interested in. Specifically, this includes: This step aims to complete the most crucial judgment step in this invention: based on the user's current interest trends. The direction of their changing interests And the user sensitivity evolution trend vector for each candidate piece of information. This paper establishes a quantifiable, direction-sensitive, and robust trend matching method to determine whether a piece of information is evolving into content that users may be interested in. Unlike the user content similarity calculation in traditional recommendation systems, this step does not aim for static matching, but emphasizes the "dynamic resonance relationship between information trends and interest trends," that is, whether the user's attention flow in the semantic space is coupled with the direction of content change during the evolution process.

[0039] To address several challenges in the trend matching process, this step proposes the following three improvement strategies: 1) Considering that the direction of user interests and the direction of content evolution are not necessarily perfectly aligned, the system introduces a dynamic direction projection mechanism, which not only measures the similarity of the current direction, but also evaluates whether the information trend vector is moving closer to the direction of interest change. 2) Considering that the evolution trend of some content may be volatile, that is, its direction is unstable at different stages, this step designed a trend stability control item to punish information with "more noise" in the trend. 3) In the platform environment, different types of information change at different rates. For example, the semantic changes of short video titles are often more dramatic, while news summaries are more stable. Therefore, a change sensitivity factor is introduced for structural compensation.

[0040] Based on this, the system uses the following trend matching function for modeling:

[0041] The first additive term is directional similarity, which measures whether the current interest direction aligns with the direction of content evolution. The second additive term introduces the user interest change vector. This measure assesses whether the content is "following the trend of user interest drift." A high value for this item indicates that the content trend is leading the way and aligns with the prediction objective. The third deduction item is a penalty for fluctuations in trend direction. The variance representing the trend direction change of this piece of information during the evolution process is defined as follows: ; in , indicating information The direction of semantic change in the Kth version This is the average direction of change across all its evolutionary stages. This item reflects the stability of content evolution by measuring whether the changes in content within the semantic space are consistent in direction, thus avoiding misjudging high-frequency fluctuations as high-trend resonance in matching scores.

[0042] The three weight terms in the above formula are as follows: To control the impact of interest drift response items on the total matching score, a setting of 0.3–0.6 is recommended. : Control the strength of the trend volatility suppression term, it is recommended to set it to 0.05–0.2; all vectors have been normalized, their directions are comparable, and the dot product result is the direction cosine value.

[0043] Taking a certain platform's public opinion monitoring scenario as an example, when users' recent interests gradually shift from "real estate market" to "green building," the system will generate a non-zero value. If a news item evolves from its initial meaning of "urban infrastructure" to "green and energy-saving renovation policy," its... Simultaneous alignment and ,but A significant increase will occur, allowing the system to determine that the information is evolving into content that users may be interested in. However, if another content direction is aligned, but the semantic fluctuations are drastic, Too high, then It will be effectively suppressed to prevent misjudgment.

[0044] This is the trend alignment score, used to quantify the "degree of trend resonance" between each candidate piece of information and the current user profile. This score will directly serve as one of the main input features of the next response probability prediction model, forming the core judgment basis in the entire system's closed-loop logic of behavior → content → trend → prediction.

[0045] Step 4: Using the response probability prediction formula, combined with factors such as information evolution stability and timeliness, perform personalized response probability estimation for each candidate message, filter candidate messages with response probabilities greater than a preset threshold, and execute corresponding personalized message pushes; specifically including: The goal of this step is to, based on the trend matching score calculation completed in the previous stage, and taking into account factors such as the stability and timeliness of information evolution, analyze each candidate piece of information. Personalized response probability estimates are performed, and actual system operations, such as push notifications, tagging, and alerts, are triggered accordingly. Unlike traditional systems that rely solely on click history or popularity models, this step takes a "trend response" perspective, emphasizing whether the information content is evolving in the direction of user attention, and whether this evolution is stable and attractive, thereby forming a truly predictive response judgment, rather than retrospective recommendations based on existing behavior.

[0046] This step includes three key variables: the first is the trend alignment score. It is achieved by step three through user interest trend vectors in the semantic space. Trends in Interest Information evolution trend vector The score is calculated using directional coupling. This score reflects whether the information is currently evolving along the semantic direction that the user is interested in, and is a core indicator used to determine whether the user is likely to respond to the content in the future.

[0047] The second variable is the uncertainty coefficient of information evolution. It retains the multi-version evolution vector sequence of information from step two. Extracted from the data. The system calculates the discreteness of the angle between each evolution vector and the overall average evolution direction to reflect whether the evolution direction of the information content in the semantic space is stable. Content with large fluctuations means that the system has difficulty predicting its subsequent state, and therefore should be penalized in the response probability.

[0048] The third variable is the information timeliness index. This refers to the time interval (in minutes or hours) between the information's publication and the current point in time. This variable is directly extracted from the content metadata of the content management system. Platform observations have revealed that users are significantly more willing to respond to newly evolved information (such as news that has just been updated or has recently added trending terms) than to content that has been published for a longer period of time. Therefore, this variable can serve as an important enhancement factor.

[0049] Based on the three input variables mentioned above, the system constructs the following response probability prediction formula to determine whether each candidate message will be responded to by the user within a future time window: ; in, It's the Sigmoid function, used to normalize input values ​​to a certain value. This generates a standard probability output. (Coefficients) and These are the weight parameters learned during model training, and are typically initialized to... , , It can also be fine-tuned based on behavioral data from different platforms. Indicates the trend matching score. The term is a penalty for fluctuations in the direction of evolution. This is an indicator of information timeliness. The structure integrates multiple factors such as structural trends, behavioral uncertainty, and information lifecycle, making it an interpretable predictive model.

[0050] For example, in a news recommendation system, a user's recent interests quickly shifted from "traditional gasoline vehicles" to "hybrid vehicles," which was achieved through... and This should be clearly reflected. If a news article evolves from "oil price fluctuations" to "hybrid subsidy policy" and then further adds "company hybrid technology release," its... It will align with user trends and generate higher Meanwhile, the news editing process was stable. The value is low and has just been updated. Small. Therefore, this information is obtained at a higher level in this formula. The system will include this information in the response set.

[0051] After the prediction is completed, the platform will adjust the settings according to the set threshold. For all Filter and keep all The candidate information is displayed, and corresponding personalized message pushes are executed.

[0052] For example, in the same recommendation system, if another user's recent interests are focused on "artificial intelligence and educational applications," the system calculates several candidate content items through trend matching. One of these, a report about "introducing AI-assisted teaching in primary and secondary schools," gradually incorporates new information such as "optimization of classroom interaction algorithms" during semantic evolution, significantly increasing its trend score and being selected by the model as content with a high response probability. Based on this, the system prioritizes pushing this report to the user's homepage recommendation feed, thus achieving a precise match between user interest trends and information evolution directions.

[0053] These operations are configured by different scenario modules. For example, in the recommendation scenario, the information can be pushed to the user's homepage or topic stream; in the risk monitoring scenario, keyword highlighting or sentiment polarity classification can be triggered; and in the digital twin supervision system, it can be automatically attached to the event evolution node and start the emergency response logic.

[0054] Please refer to Figure 2 As shown, in a second aspect of this application, a real-time information evolution prediction system based on user profiles is also proposed, comprising: The data acquisition and processing module is used to collect and sort user content interaction behaviors in real time through a fixed-length sliding window. The content involved in each behavior is input into the pre-trained BERT-base model to output semantic vectors. Combined with the weight mechanism, a user interest trend vector is constructed. Finally, through the time difference comparison mechanism, a preset time point is selected to construct the interest trend vector of the previous period and calculate the change of the current user's interest direction in the time dimension. The information evolution trend vector construction module is used to select all information items with semantic multi-version records from the current candidate information set, input them into the BERT-base model for vector encoding, obtain semantic representation, and use the user interest trend vector as the filtering function in the content trend extraction process. It scores the directional consistency of each pair of adjacent semantic change segments, and by setting a threshold, only retains all version change segments with consistency scores greater than the threshold, and sums them according to their strength to construct the semantic evolution trend vector of each candidate information that is sensitive to the current user. The personalized information trend assessment module is used to measure the similarity between a piece of information and the user's current interest direction through a trend matching function, and to assess whether the information trend vector is moving closer to the direction of interest change, and finally obtains a trend alignment score, which is used to quantify the degree of trend resonance between each candidate piece of information and the current user profile. The higher the trend alignment score, the more likely the information is evolving into content that the user is potentially interested in. The trend matching probability prediction module is used to estimate the personalized response probability of each candidate piece of information by combining the response probability prediction formula with factors such as information evolution stability and timeliness. Candidate information with a response probability greater than a preset threshold is selected and corresponding personalized message pushes are executed. The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting real-time information evolution based on user profiles, characterized in that, Includes the following steps: The system collects and sorts user content interaction behaviors in real time through a fixed-length sliding window. The content involved in each behavior is input into the pre-trained BERT-base model, which outputs a semantic vector. The system combines a weight mechanism to construct a user interest trend vector. Finally, through a time difference comparison mechanism, the system selects a preset time point to construct the interest trend vector of the previous period and calculates the change in the current user's interest direction in the time dimension. Select all information items with semantic multiple version records from the current candidate information set, input them into the BERT-base model for vector encoding, obtain semantic representation, and use the user interest trend vector as the filtering function in the content trend extraction process. Score the directional consistency of each pair of adjacent semantic change segments. By setting a threshold, only retain all version change segments with consistency scores greater than the threshold, and sum them according to their strength to construct the semantic evolution trend vector of each candidate information that is sensitive to the current user. The trend matching function measures the similarity between a piece of information and the user's current interest direction, and evaluates whether the information trend vector is moving closer to the direction of interest change. Finally, a trend alignment score is obtained, which is used to quantify the degree of trend resonance between each candidate piece of information and the current user profile. The higher the trend alignment score, the more likely the information is evolving into content that the user is potentially interested in. By using the response probability prediction formula and combining factors such as information evolution stability and timeliness, a personalized response probability estimate is performed on each candidate message. Candidate messages with response probabilities greater than a preset threshold are selected, and actual operations are triggered based on the candidate messages to execute corresponding personalized message pushes. The content interaction behavior includes content ID, behavior type, timestamp, page dwell time, and viewing ratio, wherein the behavior type includes browsing, clicking, favorite, and commenting; The directional consistency scoring includes calculating the cosine score of the angle between different versions of the semantic change vector of each piece of information content, which is used to represent the resonance intensity. In the process of calculating the cosine score, there are basic directional consistency scores and dynamic trend guidance items, which are used to measure the consistency between the content evolution direction and the current interest trend and adjust the influence ratio between the current interest and the change trend, respectively. In the response probability prediction formula, a prediction model is constructed using trend alignment score, information evolution uncertainty coefficient, and information timeliness index to output the response probability of each candidate information within a future time window; the information evolution uncertainty coefficient reflects whether the evolution direction of the information content in the semantic space is stable by calculating the discrepancy of the angle between each evolution vector and the overall average evolution direction; the information timeliness index is the time interval between the release of the information and the current time point.

2. The real-time information evolution prediction method based on user profiles according to claim 1, characterized in that, The pre-trained BERT-base model includes 12 Transformer layers with a hidden size of 768, used to concatenate the content title and body text, and normalizes the vectors during the encoding stage.

3. The method for predicting real-time information evolution based on user profiles according to claim 1, characterized in that, In the process of constructing the semantic evolution trend vector, by filtering out content change segments that do not meet the interest resonance threshold, the user interest direction and its rate of change are introduced as a filtering method for content evolution trend extraction, and finally the semantic evolution trend vector is generated.

4. The method for predicting real-time information evolution based on user profiles according to claim 1, characterized in that, The trend matching function includes a static interest alignment term, an interest drift response term, and a trend direction fluctuation penalty term. The static interest alignment term measures whether the current interest direction is consistent with the content evolution direction. The interest drift response term measures whether the content is conforming to the user's interest drift trend. The higher the value of the interest drift response term, the more predictive the content trend is, and the more consistent it is with the prediction target. The trend direction fluctuation penalty term measures whether the changes in the content in the semantic space are consistent, reflecting the stability of the content evolution and avoiding the misjudgment of high-frequency fluctuating content as high-trend resonance in the matching score.

5. The method for predicting real-time information evolution based on user profiles according to claim 1, characterized in that, The lower the trend alignment score, the less interested the user is in the information.

6. The method for predicting real-time information evolution based on user profiles according to claim 1, characterized in that, The actual operations include pushing the information to the user's homepage or topic stream in the recommendation scenario, triggering keyword highlighting or sentiment polarity classification in the risk monitoring scenario, and attaching it to the event evolution node and initiating emergency response logic in the digital twin supervision system.

7. A real-time information evolution prediction system based on user profiles, characterized in that, Including those connected sequentially: The data acquisition and processing module is used to collect and sort user content interaction behaviors in real time through a fixed-length sliding window. The content involved in each behavior is input into the pre-trained BERT-base model to output semantic vectors. Combined with the weight mechanism, a user interest trend vector is constructed. Finally, through the time difference comparison mechanism, a preset time point is selected to construct the interest trend vector of the previous period and calculate the change of the current user's interest direction in the time dimension. The information evolution trend vector construction module is used to select all information items with semantic multi-version records from the current candidate information set, input them into the BERT-base model for vector encoding, obtain semantic representation, and use the user interest trend vector as the filtering function in the content trend extraction process. It scores the directional consistency of each pair of adjacent semantic change segments, and by setting a threshold, only retains all version change segments with consistency scores greater than the threshold, and sums them according to their strength to construct the semantic evolution trend vector of each candidate information that is sensitive to the current user. The personalized information trend assessment module is used to measure the similarity between a piece of information and the user's current interest direction through a trend matching function, and to assess whether the information trend vector is moving closer to the direction of interest change, and finally obtains a trend alignment score, which is used to quantify the degree of trend resonance between each candidate piece of information and the current user profile. The higher the trend alignment score, the more likely the information is evolving into content that the user is potentially interested in. The trend matching probability prediction module is used to estimate the personalized response probability of each candidate information by combining the response probability prediction formula with the stability and timeliness of information evolution factors, and to filter candidate information with response probabilities greater than a preset threshold and execute corresponding personalized message push. Filter candidate information with a response probability greater than a preset threshold and trigger actual operation based on the candidate information to execute corresponding personalized message push; The content interaction behavior includes content ID, behavior type, timestamp, page dwell time, and viewing ratio, wherein the behavior type includes browsing, clicking, favorite, and commenting; The directional consistency scoring includes calculating the cosine score of the angle between different versions of the semantic change vector of each piece of information content, which is used to represent the resonance intensity. In the process of calculating the cosine score, there are basic directional consistency scores and dynamic trend guidance items, which are used to measure the consistency between the content evolution direction and the current interest trend and adjust the influence ratio between the current interest and the change trend, respectively. In the response probability prediction formula, a prediction model is constructed using trend alignment score, information evolution uncertainty coefficient, and information timeliness index to output the response probability of each candidate information within a future time window; the information evolution uncertainty coefficient reflects whether the evolution direction of the information content in the semantic space is stable by calculating the discrepancy of the angle between each evolution vector and the overall average evolution direction; the information timeliness index is the time interval between the release of the information and the current time point.

Images