Personalized content real-time pushing method based on user portrait

Abstract

The invention relates to the technical field of content pushing, and discloses a personalized content real-time pushing method based on a user portrait. The method comprises the steps of collecting a user behavior data set and historical interaction records, extracting real-time behavior characteristics and generating a behavior dynamic sequence; identifying behavior delay features and converting the behavior delay features into interest feedback lag; determining a dynamic influence weight in combination with an interaction relationship between the static attribute characteristics of the user and the real-time behavior characteristics; predicting the interest tendency of the user in real time by using the hysteresis and the dynamic influence weight, and generating a fluctuation deviation value; and when the deviation value exceeds a preset threshold value, triggering dynamic updating of the user portrait and adjusting a pushing strategy. According to the method, the user behavior dynamic change is captured, the behavior feedback hysteresis is considered, and the feature influence weight is dynamically adjusted, so that the user interest is accurately predicted in real time, the user portrait and the pushing strategy can be flexibly updated, the individuation and timeliness of content pushing are improved, and the real-time requirement of the user is met.

Description

Technical Field

[0001] The present invention relates to the technical field of content push, and in particular to a method for real-time push of personalized content based on user portraits. Background Art

[0002] In an era of information explosion, users are faced with a vast amount of content. How to accurately deliver content of interest to users from this vast information pool has become a crucial issue in the content services sector. Currently, most content delivery systems rely on user profiles for personalized recommendations. These systems build profiles by collecting historical user data and then match relevant content to the profiles. However, existing methods have obvious limitations. User portraits are mostly constructed based on static attributes and historical behaviors, making it difficult to capture real-time changes in user behavior. For example, a user's interests may change rapidly in different time periods and scenarios, and traditional systems often take a long time to perceive such changes, resulting in pushed content lagging behind the user's current needs; existing technologies often use fixed weights to deal with the relationship between user behavior characteristics and static attribute characteristics, ignoring the impact of the dynamic interaction between the two. For example, users with the same static attributes may show completely different interest tendencies at different stages of behavior, and the fixed weight calculation method cannot accurately reflect such dynamic changes, thus affecting the accuracy of push notifications. The delayed nature of user behavior has also not been effectively addressed. Feedback on some user behaviors isn't instantaneous, but rather lags behind. Existing methods often ignore this lag and directly predict interests based on immediate behavior, resulting in deviations between the predictions and the user's actual interests. Furthermore, the user profile update mechanism is not flexible enough. In most cases, updates rely on fixed time intervals or cumulative behavior counts, making it impossible to dynamically adjust based on real-time fluctuations in user interests, further reducing the timeliness and accuracy of push notifications. These issues collectively make it difficult for existing content push systems to meet users' growing demands for personalization and real-time performance. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for real-time push of personalized content based on user portraits to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides a method for real-time push of personalized content based on user portraits, the method comprising: Collect user behavior data sets and historical interaction records; extract user real-time behavior features from the user behavior data sets, and generate user behavior dynamic sequences based on the user real-time behavior features; Identifying multiple behavior delay features based on the changing trend of the user behavior dynamic sequence, and converting the behavior delay features into hysteresis amounts of user interest feedback; Obtaining static attribute features in the user portrait, and combining the interactive relationship between the static attribute features and the real-time behavior features to determine the dynamic influence weight of the attribute features on the behavior features; Real-time prediction of user interest tendency is performed using the hysteresis amount and the dynamic influence weight to generate a user interest fluctuation deviation value; When the user interest fluctuation deviation value exceeds the preset fluctuation threshold, the dynamic update instruction of the user portrait is triggered and the content push strategy is adjusted.

[0005] Preferably, the extracting of user real-time behavior features from the user behavior data set includes: calculating a time interval sequence of user continuous operation behaviors; Identifying peak intervals and valley intervals in the time interval sequence; and dividing user operation intensity levels based on the peak intervals and valley intervals; Extracting the depth features and dwell time features of the current session according to the user operation intensity level; The depth feature and the stay duration feature are fused into the user real-time behavior feature.

[0006] Preferably, generating a user behavior dynamic sequence based on the user's real-time behavior characteristics includes: establishing a user behavior graph network model, wherein nodes represent user operation types and edge weights represent transfer frequencies between operation types; inputting the plurality of behavior delay features into the user behavior graph network model; Update the node embedding representation through graph convolution operations to generate a sequence of operation paths with time series labels; Perform sliding window sampling on the operation path sequence and output user behavior dynamic sequence fragments.

[0007] Preferably, determining the dynamic influence weight of the attribute feature on the behavior feature includes: constructing a user preference prediction network, wherein the input layer includes a static attribute feature branch and a real-time behavior feature branch; Setting a feature cross layer in the user preference prediction network to calculate the cosine similarity matrix between static attribute features and real-time behavior features; generating a feature interaction strength coefficient according to the cosine similarity matrix; The feature interaction strength coefficient is mapped into a dynamic influence weight vector through a multi-layer perceptron.

[0008] Preferably, the user preference prediction network further performs: extracting interest tag distribution in user historical interaction records; Encoding the interest tag distribution into an interest feature vector; Performing dimensionality reduction and compression processing on the interest feature vector to generate a low-dimensional interest implicit representation; The low-dimensional interest implicit representation and the dynamic influence weight vector are weightedly concatenated to form a comprehensive preference feature.

[0009] Preferably, generating the user interest fluctuation deviation value includes: performing time slicing processing on the user behavior dynamic sequence to divide it into multiple interest analysis segments; calculating the entropy value change rate of the operation type distribution in each interest analysis segment; Extracting a sequence of entropy value change rates of the plurality of interest analysis segments; Inputting the entropy value change rate sequence and the hysteresis into a time series prediction model; The time series prediction model outputs a continuous prediction curve of the user interest fluctuation deviation value.

[0010] Preferably, the time series prediction model performs: obtaining a preset value of an interest attenuation coefficient in a user portrait; Calculate the actual decay rate of the distribution of operation types in the analysis segment of interest; generating an interest stability index according to a difference between the actual attenuation rate and a preset value of the interest attenuation coefficient; The interest stability index is compared with the user interest fluctuation deviation value prediction curve, and a deviation correction parameter is output.

[0011] Preferably, when the user interest fluctuation deviation value exceeds a preset fluctuation threshold, the method further includes: starting a real-time behavior feature recalibration module to separate noise operation events from the original user behavior data set; Updating the data expression form of the user's real-time behavior characteristics; recalculating the behavior delay characteristics and dynamic impact weights; The optimized user interest fluctuation deviation value is generated through the updated behavior delay characteristics and dynamic influence weights.

[0012] Preferably, the adjusting the content push strategy includes: generating a content feature selection instruction according to the optimized user interest fluctuation deviation value; Filtering a content vector set that meets the feature selection instruction from the content feature library; Prioritizing the content vector sets to generate a content push queue; and sending the content push queue to a content display interface of a terminal device.

[0013] Preferably, generating a content push queue includes: extracting context parameters of a user terminal device; Adjusting the display weight coefficient of the content vector according to the context parameters; Fusion of the display weight coefficient and the inherent characteristic value of the content vector to generate a push content instance with a timeliness mark; The push content instance is inserted into a designated position of a content push queue according to the timeliness mark.

[0014] Compared with the prior art, the present invention has the following beneficial effects: By collecting user behavior datasets and historical interaction records, we can fully understand user behavior trajectories and historical preferences, providing rich foundational data for subsequent feature extraction and interest prediction. Extracting real-time behavioral features from user behavior datasets and generating dynamic sequences allows us to capture user behavior beyond static, fragmented records. Instead, we can track the continuous changes in user behavior in real time, promptly reflecting the user's current behavioral tendencies and making it possible to accurately grasp the dynamics of user interests. Based on the changing trends of behavioral dynamic sequences, multiple behavioral delay features are identified and converted into the lag of user interest feedback. This effectively addresses the problem of traditional methods ignoring the lag of user behavior feedback. By considering this lag, we can more accurately understand the relationship between user behavior and interests, avoid misjudgments of interests caused by the one-sidedness of immediate behavior, and make interest predictions more tailored to the user's actual situation. By extracting static attribute features from user profiles and combining them with real-time behavioral features to determine dynamic influence weights, we break the limitations of traditional fixed weighting models. This dynamic weighting can flexibly reflect the differences in the impact of static attributes and real-time behavior on user interests in different contexts. For example, users of different age groups may have different levels of interest in the same real-time behavior. Dynamic weighting can accurately capture these differences, thereby improving the accuracy of interest prediction. By using hysteresis and dynamic influence weights to predict user interests in real time and generate a fluctuation deviation value, the system no longer relies on static insights into user interests, but instead tracks their dynamic evolution in real time. When the fluctuation deviation value exceeds a preset threshold, a dynamic update command for the user profile is triggered, adjusting the push strategy. This enables flexible updates of user profiles and real-time optimization of push strategies. This mechanism promptly responds to sudden changes in user interests, ensuring that push content remains in sync with the user's current interests and needs, and preventing push failures caused by outdated profiles or rigid strategies. This method organically combines multiple factors, including a user's real-time behavior, historical records, static attributes, and behavioral delay characteristics, to form a dynamic, closed-loop push system. Within this system, each link is interconnected and mutually influential, enabling continuous optimization of the understanding and prediction of user interests, thereby making pushed content more aligned with users' immediate needs and potential preferences, and improving user acceptance and satisfaction with pushed content. Furthermore, this method requires no complex hardware or additional user input; it can be implemented solely through in-depth mining and dynamic analysis of existing data. This method is highly practical and operational, and can be widely applied to various content push scenarios to meet the needs and characteristics of different user groups. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a diagram showing the working principle of the method for real-time push of personalized content based on user portraits according to the present invention; Figure 2 Flowchart for extracting real-time user behavior features; Figure 3 Flowchart for dynamic impact weight calculation; Figure 4 Flowchart for generating deviation values ​​of user interest fluctuations; Figure 5 Flowchart for real-time behavioral feature recalibration. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] See also Figure 1 The present invention provides a method for real-time push of personalized content based on user portraits, the method comprising: By dynamically capturing the interaction between user behavior characteristics and static attributes, content push strategies can be optimized in real time. The specific implementation process includes the following steps: collecting user behavior datasets and historical interaction records, extracting real-time user behavior characteristics, and generating dynamic sequences of user behavior. By analyzing the changing trends of dynamic sequences, behavioral delay characteristics are identified and converted into lags in user interest feedback. The static attribute features in the user profile are combined to calculate their dynamic influence weight on the behavioral characteristics. The lag and dynamic influence weight are used to predict user interest tendencies and generate an interest fluctuation deviation value. When the deviation value exceeds a preset threshold, a dynamic update of the user profile is triggered and the push strategy is adjusted.

[0018] Example 1: See Figure 2 The process of extracting real-time user behavior features starts with collecting user behavior data sets. These data include user operation records such as clicks, browsing, searching, and favorites on the platform, as well as the timestamp of each operation. The calculation of the time interval series is based on the time difference between consecutive operation behaviors, such as the interval between two clicks or the response time of a page jump. The peak interval in the time interval series represents the time period with intensive user operations, while the trough interval reflects the period with sparse operations. The identification of peaks and troughs is achieved through the sliding window statistical method. When the operation frequency in the window exceeds the preset density threshold, it is determined to be a peak interval, and when it is lower than the threshold, it is determined to be a trough interval.

[0019] The division of operation intensity levels is based on the distribution of peak and trough intervals. A high operation intensity level corresponds to user behavior within the peak interval, indicating that the user is currently active; a low operation intensity level corresponds to the trough interval, reflecting a temporary decrease in user interest or distraction. The dynamic changes in the operation intensity level are used to capture interest fluctuations in user sessions. The extraction of deep features is based on the hierarchical structure of pages visited by users in the current session, such as the depth of the jump path from the home page to the detail page, or the number of steps from search to the final click. The dwell time feature is obtained by counting the time users stay on a specific page. A longer dwell time may reflect that users pay more attention to the content.

[0020] The fusion of deep features and dwell time features uses feature concatenation to form a multi-dimensional real-time user behavior feature vector. This vector not only includes the temporal information of user actions, but also encompasses the complexity of the action path and the depth of content consumed. The dynamic update mechanism of real-time behavior features ensures that newly generated user behavior data is promptly incorporated into the feature calculation process, avoiding biased interest predictions due to data delays.

[0021] The generation of dynamic sequences of user behavior relies on the construction of a graph network model. Nodes in the graph network are defined as user operation types, such as click, play, and purchase, and the edges between nodes represent the transfer relationships between operation types. Edge weights are calculated by counting the frequency of transfers between operation types in historical data. Operation pairs with high-frequency transfers have higher edge weights. The input of behavioral delay features is achieved by marking the time delay of operation transfers, such as the time difference from click to purchase or the interval between browsing and adding to favorites. These delay features reflect the lag effect of user interest feedback. For example, a user may need to browse multiple times before deciding to purchase a certain type of product.

[0022] The graph convolution operation updates the node embedding representation, aggregating information from adjacent nodes and preserving the temporal characteristics of operation transfers. The updated node embedding contains contextual information about the operation type, for example, after clicking on a certain product, one may be inclined to browse related recommendations. The operation path sequence with time series labels is generated by traversing the high-weight edges in the graph network. Each node in the sequence is accompanied by timestamp information, forming a spatiotemporal expression of user behavior. The sliding window sampling technology segments the operation path sequence, and the window size is dynamically adjusted according to the average duration of the user session. The operation path segments within each window constitute a subset of the user behavior dynamic sequence, which is used to analyze the interest change pattern over a short period of time.

[0023] Analysis of behavioral dynamic sequence segments focuses on the stability and burstiness of action transitions. Stability is measured by recurring action patterns within the sequence, such as frequent search-click-purchase paths; burstiness is determined by newly occurring action transitions, such as a sudden increase in video playback. The dynamic update mechanism of sequence segments ensures that new user behaviors are promptly incorporated into the analysis, avoiding incomplete interest capture due to data truncation. The output format of sequence segments uses time-coded vectors, which facilitates subsequent processing by interest fluctuation analysis algorithms.

[0024] Correlation analysis between action intensity levels and behavioral dynamic sequences reveals patterns in user interest intensity. High-intensity action dynamic sequences typically feature dense action shifts and short delays, reflecting concentrated bursts of user interest. Low-intensity action dynamic sequences exhibit dispersed action patterns and long delays, indicating a period of plateauing or declining interest. This correlation provides a reference for calculating interest fluctuation deviation values ​​based on the intensity dimension.

[0025] The coordinated update mechanism between real-time behavioral features and behavioral dynamic sequences ensures temporal consistency between the two. When new user behavior data is generated, the real-time behavioral features are updated first, subsequently triggering the regeneration of the behavioral dynamic sequence. This coordinated mechanism avoids interest prediction errors caused by time asynchrony between features and sequences, improving the responsiveness of dynamic push strategies. The version control mechanism for features and sequences records the timestamp and content of each update, making it easier to track the evolution of user interests.

[0026] User behavior dynamic sequences are stored in a hierarchical structure: raw operation data is stored in the underlying database, sequence fragments are stored in the intermediate cache layer, and aggregated sequence features are stored in the higher-level analysis module. This storage structure balances data access speed with analytical depth requirements, ensuring that the real-time push system can quickly obtain necessary sequence information. Sequence fragments are indexed by user session ID and time range, supporting efficient range queries and real-time updates.

[0027] The dynamic adjustment of the behavioral delay feature is based on a statistical analysis of the time differences within sequence segments. When a change in the distribution of user action intervals is detected, the delay feature threshold automatically adapts to the new time difference range. This adaptive mechanism prevents delay feature failure due to changes in user behavior patterns. For example, when a user transitions from quick browsing to in-depth reading, the system can identify the new delay pattern and adjust the calculation of interest lag accordingly. The delay feature's incremental update algorithm only processes newly generated behavioral data, reducing computational overhead while maintaining the timeliness of the feature.

[0028] The anomaly detection module for operation path sequences identifies behavioral segments that do not conform to normal patterns, such as repeated operations within a very short period of time or extended periods of inactivity. This marking and isolation of anomalous segments prevents them from interfering with dynamic sequence analysis and triggers a reassessment process for user attention. The anomaly detection threshold is dynamically set based on the statistical distribution of historical user behavior, avoiding false positives or missed detections caused by fixed thresholds.

[0029] High-resolution timestamps ensure the precise order of events within the operational path sequence. A time synchronization protocol coordinates time records from different data sources, eliminating timing confusion caused by device clock discrepancies. Time stamp normalization converts timestamps of varying precision into a standardized format, ensuring consistency in the temporal dimension of sequence analysis.

[0030] Sliding window parameter optimization is dynamically adjusted by monitoring the coverage and real-time requirements of sequence analysis. An overly large window reduces sensitivity to changes in interest, while a too small window may introduce noise interference. The automatic window parameter adjustment algorithm balances the integrity of sequence segments with the granularity of analysis, selecting the optimal window size based on the stability of user behavior. The window sliding step size controls the degree of overlap between sequence segments; moderate overlap ensures continuous capture of changes in interest.

[0031] The graph network model's online learning mechanism continuously updates node and edge weights to adapt to changing user behavior. New operation types automatically expand network nodes, and a decay function based on historical operation transfer frequency reduces the impact of older data on edge weights. The model's incremental training algorithm processes only newly added behavior data, avoiding the computational burden of retraining on the entire data set. Network embedding dimensionality reduction technology preserves key operation transfer patterns while reducing computational complexity, improving the efficiency of dynamic sequence generation.

[0032] Example 2: See Figure 3 , the construction of the user preference prediction network starts with the input of static attribute features and real-time behavior features. Static attribute features include long-term stable features such as user demographic information, device attributes, historical preference labels, etc. These features are normalized to form a vector representation of fixed dimensions. The real-time behavior feature branch receives the dynamic behavior sequence features generated from Example 1, which include short-term behavior indicators such as the user's recent operation intensity, path depth, and stay time. The input layers of the two branches use independent feature encoders to process the dimensionality and distribution differences of different types of features respectively.

[0033] The feature cross-layer design uses cosine similarity to calculate the correlation between static attributes and real-time behavioral features. The rows of the similarity matrix represent the dimensions of static attribute features, while the columns correspond to the dimensions of real-time behavioral features. The matrix element values ​​reflect the degree of match between the two types of features along specific dimensions. Feature vectors are normalized before similarity calculation to eliminate the influence of vector length on the similarity results. Regions of high similarity in the matrix indicate a strong correlation between static attributes and current behavior, such as users of a certain age group tending to engage in certain operating patterns. The similarity matrix is ​​generated using a batch calculation method, supporting simultaneous cross-analysis of multiple user features.

[0034] The feature interaction strength coefficient is extracted from the similarity matrix through aggregation operations on the matrix elements. The coefficient calculation takes into account the varying importance of different feature dimensions, assigning higher aggregation weights to key dimensions. A dynamic adjustment mechanism for the interaction strength coefficient monitors the stability of the feature cross-combination results and automatically recalibrates the coefficient calculation parameters when significant changes in the similarity distribution are detected. The coefficient output format is a multidimensional vector, with each dimension corresponding to a quantified strength value for a specific feature cross-combination.

[0035] The multilayer perceptron receives feature interaction strength coefficients as input and learns the dynamic influence of static attributes on behavioral features through nonlinear transformations in the hidden layers. The network structure employs a residual connection design to avoid the vanishing gradient problem in deep network training. An attention mechanism is incorporated into the generation of the dynamic influence weight vector to highlight the contribution of high-interaction-strength features to the final weight distribution. The weight vector is normalized to ensure consistency across users, facilitating subsequent calculation of comprehensive preferences.

[0036] The distribution of interest tags in historical interaction records is obtained by statistically analyzing the content category preferences of users over time. The tag distribution data structure uses a probability distribution, with each tag associated with a preference strength value, reflecting the user's relative interest in that content category. The encoding process of the interest feature vector preserves the hierarchical relationship between tags, for example, associating the tags "Sports-Football" and "Sports-Basketball" to the higher-level "Sports" category. Sparse representation techniques in vectors filter out low-frequency noise tags and focus on the user's primary areas of interest.

[0037] Dimensionality reduction and compression utilizes nonlinear manifold learning methods to map high-dimensional interest feature vectors into a low-dimensional latent space. This dimensionality reduction process preserves the topological structure of the label distribution in the original feature space, ensuring that similar interests maintain proximity in the low-dimensional space. Low-dimensional implicit representations of interests are updated less frequently than real-time behavioral features and are typically recalculated on a daily or weekly basis, balancing long-term interest stability with adaptability. The implicit representation is stored using an incremental update strategy, adjusting only some dimension values ​​based on newly added interaction records to reduce computational overhead.

[0038] A gated fusion mechanism is used to concatenate the dynamic influence weight vector and the low-dimensional implicit interest representation. The fusion process learns the complementary relationship between the two types of features, controlling the information flow across each dimension through a trainable parameter matrix. The generation of comprehensive preference features incorporates feature importance ranking, prioritizing the most discriminative feature combinations for the current scenario. Feature timeliness markers record the update time of each component, providing a freshness reference for subsequent push strategies.

[0039] The user preference prediction network is trained using a two-stage strategy. In the first stage, the network's infrastructure is pretrained using historical data to learn general patterns of feature interactions and weight mapping. In the second stage, network parameters are fine-tuned using online learning based on real-time user feedback. A weighted sampling approach for training samples balances coverage across interest categories and user groups, preventing the model from favoring high-frequency behavior patterns. The network's version control mechanism retains historical parameter snapshots, enabling rapid rollback to a stable version in the event of sudden changes in interest distribution.

[0040] The monitoring system at the feature interaction layer detects changes in the distribution of the similarity matrix and identifies new correlations between user behavior patterns and static attributes. When significant new cross-patterns are detected, the recalculation of the feature interaction strength coefficient is triggered. Monitoring metrics include statistical properties such as mean shift, variance change, and singular value distribution of matrix elements, assessing the stability of feature interactions from multiple perspectives.

[0041] The adaptive learning rate of the multilayer perceptron dynamically adjusts based on the update amplitude of the weight vector. When rapid shifts in user interest are detected, the learning rate is automatically increased to accelerate model adaptation; during periods of stable interest, the learning rate is reduced to improve parameter convergence accuracy. Sparsity constraints on network parameters prevent overfitting, while regularization terms maintain the model's generalization capabilities. The batch size for parameter updates is dynamically adjusted based on system load, balancing computational efficiency and model stability.

[0042] The time-decrease function of interest tag distribution reduces the impact of older interaction records on current tag weights. The decay coefficient is set differently based on content type; for example, news content decays faster than film and television content. Cold-start tag processing assigns initial weights to newly emerging content categories, avoiding blind spots in preference prediction caused by missing historical data. Dynamic expansion of the tag hierarchy enables seamless integration of new content categories without retraining the entire tag distribution model.

[0043] The number of dimensions in the low-dimensional implicit representation of interests is automatically adjusted based on user activity. Highly active users have higher dimensions in their implicit representation to capture subtle interest differences, while low-activity users have lower dimensions to prevent overfitting. Dimension selection is based on a balance between reconstruction error and computational cost, with the optimal dimensionality configuration determined through validation set evaluation. Visual projection methods for the implicit representation help understand the distribution of users in the interest space, assisting in fine-tuning push strategies.

[0044] Version management for comprehensive preference features records content changes with each update, enabling the tracing of historical interest evolution. A feature difference comparison algorithm quantifies the distance between adjacent versions, triggering urgent adjustments to the push strategy when a sudden update is detected. The feature's distributed caching mechanism ensures fast response times for high-frequency access, employing a least recently used strategy to manage cache space.

[0045] The update review mechanism for static attribute features ensures that long-term stable user information is not erroneously modified by short-term fluctuations. The attribute change verification process checks the compatibility of new data with historical records and updates the feature representation only when it is confirmed that the attribute has indeed changed. Special handling of sensitive attributes utilizes differential privacy technology to maintain feature validity while protecting user privacy.

[0046] Conflict detection between real-time behavioral features and static attributes identifies inconsistencies between the two, such as when a user's real-time operating patterns don't match their demographic attributes. Conflict resolution strategies select a more reliable feature source based on confidence assessments or initiate manual review. Analysis of conflict logs helps identify the root causes of system feature extraction or user behavior anomalies.

[0047] The interpretable analysis tool for the user preference prediction network visualizes feature cross-paths and weight assignment logic. The analysis results are presented in interactive graphical form, demonstrating how static attributes influence the final preference predictions at different levels, enhancing model credibility and debugging efficiency. Explanatory feature extraction methods identify the original feature combinations that most influence prediction results, helping businesses understand the key factors driving user decisions.

[0048] The elastic scaling of the network architecture automatically adjusts computing resource allocation based on request load. A horizontal scaling strategy increases parallel computing instances during peak periods and reduces resources during off-peak periods to reduce costs. A dynamic resource scheduling algorithm predicts workload trends over time and ensures sufficient processing capacity is prepared in advance.

[0049] The automated feature engineering pipeline monitors changes in data distribution and triggers necessary feature re-encoding. The pipeline's anomaly detection module identifies data quality issues during feature extraction, such as an unexpected increase in missing values ​​or sudden changes in numerical ranges. Automatic remediation attempts to restore feature validity through strategies such as interpolation or rollback, and issues alerts for manual intervention in critical cases.

[0050] The confidence level of the preference prediction results quantifies the reliability of the model's predictions for the current user. Low-confidence predictions trigger additional data collection or alternative prediction strategies to prevent unreliable results from impacting push quality. The confidence level calculation considers multiple factors, including feature completeness, model fitness, and behavioral novelty.

[0051] The closed-loop learning mechanism for user feedback feeds actual interactions with pushed content back into the system as training samples. Weighted processing of feedback data distinguishes the different amounts of information in explicit and implicit feedback. A negative sample mining strategy identifies potential negative preferences for content displayed to users without interaction, balancing the ratio of positive to negative samples.

[0052] Multi-objective optimization preference fusion addresses trade-offs between different business objectives, such as conflicting demands for click-through rate and dwell time. Configurable fusion parameters allow for tailored optimization priorities based on specific scenarios, satisfying differentiated push strategy requirements. Dynamic priorities of optimization objectives automatically adjust based on changes in business metrics, maintaining an optimal balance across all performance objectives.

[0053] Preference drift detection algorithms identify gradual changes in user interests. Drift adaptation mechanisms gradually adjust model parameters, neither overreacting to short-term fluctuations nor ignoring long-term trends. Detection metrics include multi-dimensional signals such as persistent deviations in prediction error and slow changes in feature importance.

[0054] Cross-domain feature transfer learning leverages user data from other related domains to enhance preference predictions in the primary domain. The transfer process mitigates the negative transfer effects caused by domain differences and improves knowledge transfer through feature alignment and instance weighting. The source domain is selected based on a comprehensive assessment of its relevance to the target domain and its own data quality.

[0055] Example 3: See Figure 4 The calculation of user interest fluctuation deviation values ​​is based on time slicing processing technology, which divides the dynamic sequence of user behavior into multiple interest analysis segments. The length of each segment is dynamically adjusted according to the average duration of the user session and the operation density. During high-density operation periods, shorter segment lengths are used to capture detailed changes, while during low-density periods, the segment length is appropriately extended to avoid data sparsity. The establishment of segment boundaries not only considers the time interval, but also combines the natural transition points of the operation type, such as the time node from browsing behavior to purchasing behavior. The segmentation process retains the temporal relationship of the original sequence and ensures the continuity between adjacent segments.

[0056] The entropy calculation of the operation type distribution reflects the concentration of user interests. For each interest analysis segment, the frequency of occurrence of different operation types is counted and normalized into a probability distribution. The entropy calculation uses the following formula:

[0057] in, Indicates time The entropy value of is the total number of operation types, It is The probability of a similar action occurring in the current segment. The fluctuations in entropy reveal the stability of user interests. Low entropy indicates a concentration of action types, reflecting a clear focus; high entropy indicates a dispersion of actions, indicating an exploratory or free-flowing interest. The entropy change rate is calculated by comparing the entropy differences between adjacent segments. Positive values ​​indicate a trend toward dispersed interest, while negative values ​​reflect a gradual focus.

[0058] The construction of the entropy rate of change series takes into account the effect of time decay, giving higher weight to recent changes. Smoothing the series eliminates noise caused by short-term fluctuations while preserving trends. The introduction of a lag adjusts the temporal alignment of the entropy rate of change, compensating for the inherent delay between behavior onset and interest feedback. The adjusted rate of change series is input into a time series forecasting model, which employs an attention mechanism to enhance its ability to capture key turning points.

[0059] The architecture of the time series forecasting model consists of an encoder and a decoder. The encoder converts the entropy rate of change sequence into a latent state representation, capturing long-term dependencies within the sequence. The decoder generates a curve predicting interest fluctuations for future periods based on the latent state. Each point on the curve corresponds to the expected deviation value at a specific moment. The model is trained using a multi-step forecasting strategy, forcing the learning of long-term dynamic patterns in the sequence rather than simple one-step predictions. The confidence intervals of the forecast curves are estimated using Monte Carlo dropout techniques to reflect the uncertainty of the forecast results.

[0060] The preset value for the interest decay coefficient is derived from long-term behavioral statistics within the user profile and represents the natural rate of decline in interest for that user. The actual decay rate is calculated by fitting the entropy trends of multiple recent analysis segments. The difference between the two values ​​generates the interest stability index. A positive value indicates faster-than-expected interest decay, while a negative value indicates above-average interest persistence. The index is calculated using a sliding window approach, with the window size adaptively adjusted to match the user's behavioral rhythm.

[0061] The bias correction parameter generation process compares the accuracy of the prediction curve against the stability indicator. When systematic deviations are detected, the correction parameters adjust the slope and amplitude of the prediction curve to better align with the observed patterns of interest. This correction process employs a gradual adjustment strategy to avoid oscillations in the prediction caused by a single, drastic correction. The corrected prediction curve is output as a continuous function of time, supporting interpolation of bias values ​​at any given moment.

[0062] The reassessment mechanism for interest analysis segments is triggered when an unusual entropy change is detected. This determination is based on the statistical characteristics of the historical segment distribution; changes exceeding three standard deviations are considered abnormal. The reassessment process temporarily excludes the influence of unusual segments and re-incorporates them into the analysis once sufficient data has accumulated. The segment's credibility score guides the subsequent input weighting of the prediction model, and the influence of low-credibility segments is appropriately suppressed.

[0063] Hierarchical clustering of operation types optimizes the granularity of entropy calculation. Similar operations are aggregated into higher-level categories, avoiding entropy fluctuations and distortions caused by overly detailed classification. Clustering criteria not only consider superficial similarities in operation names but also analyze transition probabilities and contextual relationships between operations. A dynamic clustering algorithm automatically expands the category system based on emerging operation types, ensuring timely classification.

[0064] The online update strategy for time series forecasting models balances the contribution of new and old data. Recent data is given a higher weight, while some historical data is retained to maintain the model's generalization capabilities. The update frequency is automatically adjusted based on the forecast error, with the update interval shortened as the error increases. A version rollback mechanism restores the model to a stable version when performance degrades significantly, ensuring the continuity of forecasting services.

[0065] Normalization of the entropy change rate eliminates absolute differences between users. The normalization baseline uses a moving average of the user's historical entropy values, highlighting relative changes rather than absolute levels. Cross-user comparisons use percentile transformation to map the raw entropy values ​​to a uniform comparison scale. The normalization parameters are regularly updated and adapted.

[0066] Example 4: See Figure 5 The trigger condition for the real-time behavioral feature recalibration module is that the deviation value of the user's interest fluctuation exceeds the preset threshold. When the system detects that the deviation value of a user is higher than the threshold for three consecutive analysis cycles, the recalibration process is initiated. The process first separates noise operation events from the original user behavior dataset. For example, a user usually browses clothing products on an e-commerce platform, but suddenly clicks on digital products multiple times in a row, and each click lasts less than 1 second. Such abnormal operations are marked as potential noise events and are screened according to the following rules: the operation interval is abnormally shorter than the historical average level, the operation type and user profile match too poorly, and incoherent behavior patterns are suddenly inserted into the operation sequence.

[0067] The identification of noise events uses a multi-dimensional cross-validation method. The following table shows an example of the analysis of user behavior data, including the original operation records and noise determination results: Table 1: User behavior data analysis example data

[0068] Timestamp Operation Type Page Category Dwell time (seconds) Interval between adjacent operations (seconds) Matching with historical preferences Noise Marker 2025-07-3110:05:31 Click Men's 12 8 0.92 normal 2025-07-3110:05:43 Browse Men's 23 0 0.89 normal 2025-07-3110:06:06 Click cell phone 0.8 5 0.12 noise 2025-07-3110:06:11 Click laptop 0.6 0 0.15 noise 2025-07-3110:06:12 return front page 2 1 0.45 To be determined 2025-07-3110:06:14 Click Men's 15 2 0.91 normal The noise determination in the table integrates multiple dimensions, including action type, dwell time, interval time, and historical preference match. Actions marked "pending" require further contextual analysis before a final classification is determined. After a noise event is isolated, the system retains a complete copy of the original data for subsequent audits and temporarily excludes the impact of this data from the processing flow.

[0069] Updating the data representation of real-time user behavior features involves a redistribution of feature weights. For example, in a reading app, users initially focused on technology articles, resulting in a "technology" label weight of 0.85 in the feature vector. A sudden, rapid influx of entertainment articles caused the weight of the "entertainment" label in the original feature vector to rise abnormally to 0.6. The recalibration process reduces the contribution of this transient abnormal behavior, returning the weight of the "entertainment" label in the adjusted feature vector to a reasonable range of approximately 0.2. The amplitude of feature expression updates is controlled incrementally, with each adjustment not exceeding 30% of the previous version's feature value to avoid system oscillations caused by a single, drastic change.

[0070] The recalculation of behavioral delay features focuses on changes in temporal patterns in normal operation sequences. For example, for users of a certain video platform, historical data showed an average delay of 45 minutes from browsing to adding a favorite. However, recently, a large number of browsing to adding a favorite has occurred within 3 minutes. After recalibration, the system recognizes that such rapid collections may be due to improvements in the quality of content recommendations rather than sudden changes in user interest. Therefore, the calculation of delay features is divided into two modes: the regular browsing-adding to favorites path retains the original delay calculation, while a separate delay feature sub-model is established for the fast adding to favorites path. This layered processing more accurately reflects the speed of interest feedback under different behavioral patterns.

[0071] Dynamic influence weights are optimized by comparing the interaction strength between static attributes and calibrated behaviors. For example, for the age attribute, the dynamic weight for "fashion" content was originally 0.7 for users aged 25-30. However, a brief, noisy behavior by a particular user caused this weight to fluctuate abnormally to 0.3. The recalibration process analyzes the similarity of this user's calibrated behavior with other users of the same age group, restoring the weight to a reasonable range of 0.65. The iterative weight optimization process uses small steps and multiple updates. After each iteration, coordination with other relevant features is checked to ensure the rationality of the overall weight distribution.

[0072] The optimized mechanism for generating deviation values ​​for user interest fluctuations incorporates a robustness check. The system compares the results of three calculation paths: one based on complete raw data, one that excludes all suspected noise, and one that partially excludes noise. When the deviation values ​​of the three paths differ beyond the permitted range, a manual review process is triggered. Under normal circumstances, the optimized deviation value is the median of the three paths, balancing the stringency of noise filtering with the preservation of data integrity. The final deviation value output is accompanied by a confidence score, reflecting the reliability of the current calculation result.

[0073] An example of the logic behind generating content feature selection instructions: After recalibration, a music user's interest fluctuation deviation value shows that their preference for "electronic music" has dropped from 0.8 to 0.5, while their interest in "jazz" has increased from 0.3 to 0.6. Based on this, the system generates an instruction requesting the content library to return content items where the jazz feature weight is higher than that of electronic music. Detailed parameters for the instruction include: primary tag weight range, secondary tag exclusion list, content freshness preference, etc. The instruction priority setting takes into account the magnitude of the deviation value change, with interest points with large fluctuations receiving higher selection priority.

[0074] The content presentation interface on the terminal device handles the final presentation of the push queue. The interface optimizes content layout based on device type and screen size, prioritizing concise information on mobile devices and providing richer, supplementary content on tablets. The presentation logic considers the user's current operating environment, for example, automatically lowering the priority of video content on mobile networks and restoring full recommendation weighting on WiFi. The interface's feedback collection mechanism records discrepancies between actual user interactions and the expected push notifications. This data is fed back into the evaluation phase of the recalibration module.

[0075] The execution of dynamic update commands utilizes a transactional mechanism to ensure data consistency. Updates must either complete successfully or be completely rolled back to the previous stable state. The version control system records a detailed log of each update, including the triggering cause, modified content, and execution results. Changes to key parameters require double verification, especially adjustments to key user interests, which must undergo additional consistency checks.

[0076] The noise separation module's self-learning function gradually refines filtering rules. The system regularly analyzes the subsequent development of actions marked as noise to confirm whether they are truly abnormal. Misidentifications trigger adjustments to filtering rule parameters. For example, if a short but genuine exploration of new interests is frequently misclassified as noise, the system automatically relaxes the threshold for such actions. The learning process is accelerated to prevent over-adjustment of long-standing empirical rules due to short-term pattern changes.

[0077] A framework for AB testing content push strategies verifies the effectiveness of adjustments. User traffic is randomly assigned to different strategy groups, and new strategies are fully rolled out only when they outperform the baseline. Testing metrics include not only traditional indicators such as click-through rate, but also comprehensive results such as long-term user retention and satisfaction surveys. A phased release mechanism allows for the gradual expansion of new strategies' coverage while closely monitoring trends across various metrics.

[0078] The exception recovery mechanism automatically activates when the system detects a decline in push performance. The recovery process first rolls back the push strategy of the previous stable version while conducting a thorough analysis of the root cause of the issue. Common recovery actions include resetting some feature weights, clearing potentially outdated cached data, and strengthening noise filtering. User feedback during the recovery process is given greater weight, accelerating the system's adaptive adjustments.

[0079] Prioritizing explicit user feedback overrides automatic push decisions. When users proactively bookmark, rate, or complain about content, this feedback immediately influences adjustments to the current push queue, eliminating the need to wait for regular recalibration cycles. Explicit feedback credibility assessment weights different types of feedback, such as a five-star rating being more reliable than a simple click. Semantic analysis of feedback content extracts the user's direct intent, preventing misunderstandings caused by simple statistics.

[0080] Example 5: The generation process of the content push queue comprehensively considers the contextual environment parameters of the terminal device, which include but are not limited to device type, network status, geographic location, current time, battery level, and ambient light. The device type information determines the layout and interaction mode of content display. For example, smartphones use a single-column waterfall display, while smart TVs use a grid layout. Network status parameters are divided into three basic modes: mobile network, WiFi, and offline. There are significant differences in content loading strategies under different states. In a mobile network environment, text and low-resolution image content are pushed first, while in a WiFi environment, the default weight of high-definition images and videos is restored. Geographical location information is used to adjust the priority of region-related content. For example, the display weight of local news and life services will change dynamically according to the user's location.

[0081] The collection of contextual environmental parameters is obtained in real time through the device API, and the sampling frequency is dynamically adjusted according to the parameter type. Parameters that change rapidly, such as network status and battery level, are sampled at high frequencies, while stable parameters, such as device type and screen size, are only collected once during initialization. Parameter normalization eliminates the impact of different dimensions and maps various parameters to a unified scoring scale. Timeliness management of environmental parameters sets a reasonable expiration time, and parameters that exceed the validity period need to be re-collected and verified. Correlation analysis between parameters identifies potential combined impacts, such as the need for specially optimized content strategies when low battery and weak network signals occur simultaneously.

[0082] The calculation of the display weight coefficient is based on the degree of match between context parameters and content features. Each content vector contains multiple display-related features, such as text length, image ratio, video length, and interaction complexity. The coefficient calculation process first evaluates the compatibility of content features with the current device environment. For example, the display weight of large text content on smartwatches will be appropriately reduced. Secondly, the impact of network status on content loading speed is considered. Content with high bandwidth requirements is downgraded in mobile network environments. Finally, the display format is adjusted based on the ambient light and time information. The content contrast and brightness parameters in night mode are automatically optimized.

[0083] The process of extracting the intrinsic eigenvalues ​​of content vectors takes into account the attributes and historical performance of the content itself. Text content is analyzed for keyword density and sentiment, image content identifies key visual elements and color distribution, and video content extracts keyframe features and rhythm changes. Historical performance data includes metrics such as the average click-through rate, completion rate, and number of shares for similar content. The dynamic update mechanism for intrinsic eigenvalues ​​tracks changing trends throughout the content lifecycle. The initial value of newly released content is set based on the average performance of similar content and is gradually adjusted to actual data as exposure increases. Smoothing of eigenvalues ​​avoids drastic changes in display weight due to short-term fluctuations.

[0084] The process of generating push content instances with time-sensitive tags integrates display weight coefficients and inherent characteristic values. Time-sensitive tags reflect the optimal display period for content, such as limited-time promotional content with clear start and end times, or news and information with recommended prime reading times. Instance generation utilizes a templated approach, applying preset display templates to different content types while retaining a certain degree of adaptive adjustment. Template parameters include details such as title font size, image cropping ratio, and interactive button placement, all of which are automatically optimized based on the device environment. The instance verification process checks whether the generated results meet display specifications to avoid issues such as layout errors and content truncation.

[0085] The insertion position of pushed content instances in the queue is determined by a combination of timeliness tags and overall priority. Content with strict time limits is given queue-jumping privileges, and instances approaching their deadlines are automatically ranked higher. The ranking of regular content takes into account multiple factors, including user interest matching, content freshness, and a balance of content diversity. The queue's real-time adjustment mechanism responds to changes in environmental parameters. For example, when the network switches from mobile data to WiFi, it automatically moves video content forward. A hierarchical strategy is used to resolve insertion conflicts. Content of the same level is sorted chronologically, and content of different levels is allowed to preempt high-priority instances.

[0086] The content display interface on the terminal device handles the final rendering of the queue instances. The interface optimizes resource loading order based on device performance, prioritizing first-screen content and utilizing lazy loading for non-visual areas. Adaptive adjustments during the rendering process take into account actual display performance. When it detects that certain content elements are unable to display properly, a degraded display solution is automatically enabled. The interface's interaction tracking module records user behavior data such as scroll depth, dwell time, and action path. This data is fed back to the queue generation process to optimize subsequent push notifications.

[0087] Conflict resolution for time-sensitive tags addresses overlapping time requirements for multiple content instances. A conflict detection algorithm identifies overlapping time windows and prioritizes them based on business rules. For example, promotional content typically uses a first-come, first-served basis for time conflicts, while urgent notifications can interrupt regular content. Conflict resolution records detailed logs for subsequent analysis and rule optimization.

[0088] Adaptive device performance adjustment optimizes content performance on different hardware. Low-performance devices receive simplified content instances, disabling unnecessary animations and background loading. The performance monitoring module tracks content rendering frame rate and response latency, automatically triggering optimization processes when performance bottlenecks are detected. Long-term performance statistics build a profile of device capabilities, providing a predictive basis for subsequent push notifications.

[0089] The display of multimodal content coordinates the presentation order of mixed content types. For example, the alternation of mixed text and image content with pure video content must take into account the user's reading rhythm and attention shifts. The coordination algorithm analyzes the user's historical consumption patterns of different content types to find the optimal display intervals and transition methods. The associated push of multimodal content appropriately combines semantically related but different content formats to provide a more complete information acquisition experience.

[0090] Content display optimization for ambient light adaptation takes into account the impact of circadian rhythms on user preferences. Day mode uses bright, high-contrast tones, while night mode switches to a dark theme with low blue light. Light sensor data is integrated with time information to avoid unnecessary theme switching caused by brief environmental changes. Adaptive de-jitter processing eliminates interface flickering caused by rapid light fluctuations.

[0091] Battery-aware content policy adjustments extend device life. During low battery conditions, background data prefetching and content updates are reduced, and non-essential visual effects are suspended. A battery prediction model estimates remaining time based on current usage patterns and adjusts the aggressiveness of content push accordingly. A compensation mechanism gradually restores full functionality after battery is restored, avoiding sudden surges in resource demand.

[0092] Geographically relevant content push takes into account user movement. Local content in your usual area remains consistently displayed, while destination-related content is pushed to temporary locations during travel. Trajectory prediction algorithms analyze the user's speed and direction of movement to predict services in areas of potential interest. Geofencing technology triggers content switching in specific areas, for example, automatically prioritizing merchant promotions upon entering a shopping mall.

[0093] Cross-device content synchronization ensures a consistent user experience. Users can continue browsing content halfway through on their mobile device from the same point on their tablet. The synchronization mechanism identifies content breakpoints and records user progress and operational status. When content is modified simultaneously on multiple devices, the conflict resolution strategy generally uses the data from the device with the latest update as the default. Critical operations require explicit confirmation.

[0094] Personalized recognition for multi-account devices seamlessly handles family-shared device scenarios. User identification technology distinguishes different users on the same device, switching content push strategies based on login status or behavioral characteristics. Limited push notifications in guest mode provide universal content, protecting the primary user's privacy preferences. Intelligent recognition of account associations establishes content recommendations for family members with user permission, identifying potential shared interests.

[0095] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0096] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for real-time push of personalized content based on user portraits, characterized in that: The following steps are involved: Collect user behavior data sets and historical interaction records; extract user real-time behavior features from the user behavior data sets, and generate user behavior dynamic sequences based on the user real-time behavior features; Identifying multiple behavior delay features based on the changing trend of the user behavior dynamic sequence, and converting the behavior delay features into hysteresis amounts of user interest feedback; Obtaining static attribute features in the user portrait, combining the interactive relationship between the static attribute features and the real-time behavior features, and determining the dynamic influence weight of the attribute features on the behavior features; Performing real-time prediction of user interest tendency by using the hysteresis amount and the dynamic influence weight to generate a user interest fluctuation deviation value; When the user interest fluctuation deviation value exceeds the preset fluctuation threshold, the dynamic update instruction of the user portrait is triggered and the content push strategy is adjusted.

2. The method for real-time personalized content push based on user portraits according to claim 1, characterized in that: Extracting the user's real-time behavior features from the user behavior dataset includes: Calculate the time interval sequence of the user's continuous operation behavior; Identifying peak intervals and valley intervals in the time interval sequence; and dividing user operation intensity levels based on the peak intervals and valley intervals; Extracting the depth features and dwell time features of the current session according to the user operation intensity level; The depth feature and the stay duration feature are fused into the user real-time behavior feature.

3. The method for real-time personalized content push based on user portraits according to claim 1, characterized in that: Generating a user behavior dynamic sequence based on the user's real-time behavior characteristics includes: Establish a user behavior graph network model, where nodes represent user operation types and edge weights represent the frequency of transitions between operation types; inputting the plurality of behavior delay features into the user behavior graph network model; Update the node embedding representation through graph convolution operations to generate a sequence of operation paths with time series labels; Perform sliding window sampling on the operation path sequence and output user behavior dynamic sequence fragments.

4. The method for real-time personalized content push based on user portraits according to claim 1, characterized in that: Determining the dynamic influence weight of the attribute feature on the behavior feature includes: Build a user preference prediction network, where the input layer contains static attribute feature branches and real-time behavior feature branches; Setting a feature cross layer in the user preference prediction network to calculate the cosine similarity matrix of static attribute features and real-time behavior features; generating a feature interaction strength coefficient according to the cosine similarity matrix; The feature interaction strength coefficient is mapped into a dynamic influence weight vector through a multi-layer perceptron.

5. The method for real-time personalized content push based on user portrait according to claim 4, characterized in that: The user preference prediction network also performs: Extract the distribution of interest tags from user historical interaction records; Encoding the interest tag distribution into an interest feature vector; Performing dimensionality reduction and compression processing on the interest feature vector to generate a low-dimensional interest implicit representation; The low-dimensional interest implicit representation and the dynamic influence weight vector are weightedly concatenated to form a comprehensive preference feature.

6. The method for real-time personalized content push based on user portraits according to claim 1, characterized in that: Generating the user interest fluctuation deviation value includes: Perform time-slicing processing on the user behavior dynamic sequence to divide it into multiple interest analysis segments; calculate the entropy change rate of the operation type distribution within each interest analysis segment; Extracting a sequence of entropy value change rates of the plurality of interest analysis segments; Inputting the entropy value change rate sequence and the hysteresis into a time series prediction model; The time series prediction model outputs a continuous prediction curve of the user interest fluctuation deviation value.

7. The method for real-time personalized content push based on user portraits according to claim 6, characterized in that: The time series forecasting model performs: Get the preset value of the interest attenuation coefficient in the user portrait; Calculate the actual decay rate of the distribution of operation types in the analysis segment of interest; generating an interest stability index according to a difference between the actual attenuation rate and a preset value of the interest attenuation coefficient; The interest stability index is compared with the user interest fluctuation deviation value prediction curve, and a deviation correction parameter is output.

8. The method for real-time personalized content push based on user portraits according to claim 1, characterized in that: When the user interest fluctuation deviation value exceeds a preset fluctuation threshold, the method further includes: Launch the real-time behavior feature recalibration module to separate noise operation events from the original user behavior dataset; Updating the data expression form of the user's real-time behavior characteristics; recalculating the behavior delay characteristics and dynamic impact weights; The optimized user interest fluctuation deviation value is generated through the updated behavior delay characteristics and dynamic influence weights.

9. The method for real-time personalized content push based on user portraits according to claim 1, characterized in that: The adjustment of the content push strategy includes: Generate content feature selection instructions based on the optimized user interest fluctuation deviation value; Filtering a content vector set that meets the feature selection instruction from the content feature library; Prioritizing the content vector sets to generate a content push queue; and sending the content push queue to a content display interface of a terminal device.

10. The method for real-time personalized content push based on user portrait according to claim 9, characterized in that: Generating a content push queue includes: Extracting context parameters of the user terminal device; Adjusting the display weight coefficient of the content vector according to the context parameters; Fusion of the display weight coefficient and the inherent characteristic value of the content vector to generate a push content instance with a timeliness mark; The push content instance is inserted into a designated position of a content push queue according to the timeliness mark.

Images