A personalized content real-time pushing method based on user portrait

Abstract

The application 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 following steps: collecting a user behavior data set and historical interaction records, extracting real-time behavior features and generating a behavior dynamic sequence; identifying behavior delay features and converting the behavior delay features into interest feedback lag quantities; combining the interaction relationship between user static attribute features and real-time behavior features to determine dynamic influence weights; using the lag quantities and the dynamic influence weights to real-time predict user interest tendencies, and generating fluctuation deviation values; when the deviation values exceed preset threshold values, triggering dynamic updating of the user portrait and adjusting a pushing strategy. The method can realize accurate real-time prediction of user interests by capturing dynamic changes in user behaviors, considering behavior feedback hysteresis, and dynamically adjusting feature influence weights, can flexibly update the user portrait and the pushing strategy, can improve the personalization and timeliness of content pushing, and can meet real-time demands of users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of content pushing, in particular to a personalized content real-time pushing method based on user portrait. BACKGROUND

[0002] In the era of information explosion, users are faced with a vast amount of content information. How to accurately push the content of interest to users from the huge information base has become an important issue in the field of content service. At present, most content pushing systems rely on user portrait for personalized recommendation, which constructs the portrait by collecting the historical data of users, and then matches the corresponding content based on the portrait.

[0003] However, the existing method has obvious limitations. The user portrait is mostly constructed based on static attributes and historical behaviors, which is difficult to capture the real-time behavior changes of users. For example, the interests of users may change rapidly at different time periods or in different scenarios, and the traditional system often needs a long time to perceive such changes, resulting in that the pushed content lags behind the current needs of users. The existing technology mostly uses a fixed weight method to handle the relationship between user behavior features and static attribute features, ignoring the influence of dynamic interaction between them. For example, users with the same static attribute may show completely different interest tendencies at different behavior stages, and the fixed weight calculation method cannot accurately reflect such dynamic changes, thereby affecting the accuracy of pushing.

[0004] The delay characteristics existing in user behavior are also not effectively handled. The feedback of some user behaviors is not instant, but has a certain time lag. The existing method often ignores this lag and directly predicts the interest based on the instant behavior, resulting in a deviation between the prediction result and the actual interest of users. At the same time, the updating mechanism of user portrait is not flexible enough. In most cases, it relies on fixed time intervals or cumulative behavior quantities for updating, and cannot dynamically adjust according to the real-time fluctuations of user interest, further reducing the timeliness and accuracy of pushing. These problems together make it difficult for the existing content pushing system to meet the growing needs of users in terms of personalization and real-time performance. SUMMARY

[0005] The present application aims to provide a personalized content real-time pushing method based on user portrait to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a personalized content real-time pushing method based on user portrait, which comprises:

[0007] Collecting a user behavior data set and historical interaction records; extracting user real-time behavior features from the user behavior data set, and generating a user behavior dynamic sequence according to the user real-time behavior features;

[0008] According to the change trend of the user behavior dynamic sequence, a plurality of behavior delay features are identified, and the behavior delay features are converted into a lag amount of user interest feedback;

[0009] Static attribute features in the user portrait are acquired, and a dynamic influence weight of the attribute features on the behavior features is determined according to an interaction relationship between the static attribute features and the real-time behavior features;

[0010] The user interest trend is predicted in real time through the lag amount and the dynamic influence weight, and a user interest fluctuation deviation value is generated;

[0011] When the user interest fluctuation deviation value exceeds a preset fluctuation threshold, a dynamic update instruction of the user portrait is triggered, and a content pushing strategy is adjusted.

[0012] Preferably, the extracting of the user real-time behavior features from the user behavior data set comprises: calculating a time interval sequence of continuous operation behaviors of the user;

[0013] Peak intervals and trough intervals in the time interval sequence are identified, and a user operation intensity level is divided based on the peak intervals and the trough intervals;

[0014] Depth features and stay duration features of a current session are extracted according to the user operation intensity level;

[0015] The depth features and the stay duration features are fused into the user real-time behavior features.

[0016] Preferably, the generating of the user behavior dynamic sequence according to the user real-time behavior features comprises: establishing a user behavior graph network model, wherein a node represents a user operation type, and an edge weight represents a transition frequency between operation types;

[0017] The plurality of behavior delay features are input into the user behavior graph network model;

[0018] Node embedding representations are updated through graph convolution operation, and an operation path sequence with time sequence labels is generated;

[0019] The operation path sequence is sampled through a sliding window, and a user behavior dynamic sequence segment is output.

[0020] Preferably, the determining of the dynamic influence weight of the attribute features on the behavior features comprises: constructing a user preference prediction network, and an input layer comprises a static attribute feature branch and a real-time behavior feature branch;

[0021] A feature cross layer is set in the user preference prediction network, and a cosine similarity matrix of the static attribute features and the real-time behavior features is calculated;

[0022] generate a feature interaction intensity coefficient according to the cosine similarity matrix;

[0023] map the feature interaction intensity coefficient to a dynamic influence weight vector through a multilayer perceptron.

[0024] Preferably, the user preference prediction network further performs: extracting an interest label distribution in a user historical interaction record;

[0025] encoding the interest label distribution into an interest feature vector;

[0026] dimensionality reduction compression processing is performed on the interest feature vector to generate a low-dimensional interest implicit representation;

[0027] The low-dimensional interest implicit representation is weighted and spliced with the dynamic influence weight vector to form a comprehensive preference feature.

[0028] Preferably, the generating a user interest fluctuation bias value comprises: performing time slicing processing on a user behavior dynamic sequence to divide a plurality of interest analysis segments; calculating an entropy value change rate of operation type distribution in each interest analysis segment;

[0029] extracting an entropy value change rate sequence of the plurality of interest analysis segments;

[0030] inputting the entropy value change rate sequence and the lag quantity into a time series prediction model;

[0031] outputting a continuous prediction curve of the user interest fluctuation bias value through the time series prediction model.

[0032] Preferably, the time series prediction model performs: obtaining an interest decay coefficient preset value in a user portrait;

[0033] calculating an actual decay rate of operation type distribution in the interest analysis segment;

[0034] generating an interest stability index according to a difference between the actual decay rate and the interest decay coefficient preset value;

[0035] comparing the interest stability index with the user interest fluctuation bias value prediction curve to output a bias correction parameter.

[0036] Preferably, when the user interest fluctuation bias value exceeds a preset fluctuation threshold, the method further comprises: starting a real-time behavior feature recalibration module to separate noise operation events from an original user behavior data set;

[0037] updating a data expression form of the user real-time behavior feature; recalculating the behavior delay feature and the dynamic influence weight;

[0038] The optimized user interest fluctuation bias value is generated through the updated behavior delay feature and dynamic influence weight.

[0039] Preferably, the adjusting the content pushing strategy comprises: generating a content feature selection instruction according to the optimized user interest fluctuation bias value.

[0040] Filtering a content vector set meeting the feature selection instruction from a content feature library;

[0041] Prioritizing the content vector set to generate a content pushing queue; and sending the content pushing queue to a content display interface of a terminal device.

[0042] Preferably, the generating the content pushing queue comprises: extracting a context environment parameter of a user terminal device.

[0043] Adjusting a display weight coefficient of the content vector according to the context environment parameter;

[0044] Fusing the display weight coefficient and an inherent feature value of the content vector to generate a pushing content instance with a time-effectiveness label;

[0045] Inserting the pushing content instance into a specified position of the content pushing queue according to the time-effectiveness label.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] By collecting a user behavior data set and historical interaction records, the behavior trajectory and historical preference of the user can be comprehensively mastered, and rich basic data can be provided for subsequent feature extraction and interest prediction. Real-time behavior features are extracted from the user behavior data set and dynamic sequences are generated, so that the capture of user behavior is no longer limited to static and fragmented records, but can track the continuous change of user behavior in real time, timely reflect the current behavior tendency of the user, and provide the possibility of accurately grasping the user interest dynamics.

[0048] According to the change trend of the behavior dynamic sequence, a plurality of behavior delay features are identified and converted into a user interest feedback lag, effectively solving the problem of ignoring the user behavior feedback lag in the traditional method. By considering this lag, the relationship between user behavior and interest can be more accurately understood, the interest misjudgment caused by the one-sidedness of the instant behavior can be avoided, and the interest prediction is more in line with the actual situation of the user.

[0049] The static attribute features in the user portrait are acquired, and a dynamic influence weight is determined in combination with the interaction relationship between the static attribute features and real-time behavior features, thereby breaking the limitation of the traditional fixed weight mode. The dynamic weight setting can flexibly reflect the difference in the influence of static attributes and real-time behaviors on user interest in different situations, for example, users of different ages may have different interest depths under the same real-time behavior, and the dynamic weight can accurately capture such differences, thereby improving the accuracy of interest prediction.

[0050] The user interest tendency is predicted in real time through the lag amount and the dynamic influence weight, and a fluctuation deviation value is generated, so that the understanding of the user interest is no longer a static result, but can track the dynamic change process in real time. When the fluctuation deviation value exceeds a preset threshold, a user portrait dynamic update instruction is triggered and the push strategy is adjusted, realizing flexible updating of the user portrait and real-time optimization of the push strategy. This mechanism can timely respond to the mutation of user interest, ensure that the push content is always synchronized with the current interest demand of the user, and avoid push failure caused by outdated portrait or rigid strategy.

[0051] The method combines the real-time behavior, historical records, static attributes, and behavior delay features of the user, and forms a dynamic closed-loop push system. In this system, each link is interrelated and interrelated, and can continuously optimize the understanding and prediction of user interest, so that the push content is more in line with the immediate needs and potential preferences of the user, and the acceptance and satisfaction of the user to the push content are improved. In addition, the method does not need to rely on complex hardware devices or additional user input, and can be realized only by deep mining and dynamic analysis of existing data, which has strong practicability and operability, and can be widely applied to various content push scenes, and is suitable for the demand characteristics of different user groups. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The working principle diagram of the personalized content real-time push method based on the user portrait described in the application;

[0053] Figure 2 The flowchart for extracting real-time behavior features of the user;

[0054] Figure 3 The flowchart for calculating the dynamic influence weight;

[0055] Figure 4 The flowchart for generating the user interest fluctuation deviation value;

[0056] Figure 5 The flowchart for recalibrating the real-time behavior features. DETAILED DESCRIPTION

[0057] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0058] With reference to Figure 1 The present application provides a personalized content real-time pushing method based on user portrait, which comprises the following steps:

[0059] By dynamically capturing the interaction relationship between user behavior characteristics and static attributes, the real-time optimization of content pushing strategy is realized. The specific implementation process comprises the following steps: collecting user behavior data set and historical interaction record, extracting user real-time behavior characteristics and generating user behavior dynamic sequence. By analyzing the change trend of the behavior dynamic sequence, the behavior delay characteristics are identified and converted into the lag amount of user interest feedback. Combined with the static attribute characteristics in the user portrait, the dynamic influence weight of the behavior characteristics is calculated. The user interest tendency is predicted by using the lag amount and the dynamic influence weight, and the interest fluctuation deviation value is generated. When the deviation value exceeds the preset threshold, the user portrait dynamic update is triggered and the pushing strategy is adjusted.

[0060] Embodiment 1: With reference to Figure 2 The extraction process of user real-time behavior characteristics starts from collecting user behavior data set, which includes user's operation records such as clicking, browsing, searching, collecting and so on on the platform, as well as the time stamp of each operation. The calculation of time interval sequence is based on the time difference between continuous operation behaviors, such as the interval length between two clicks or the response time of page jump. The peak interval in the time interval sequence represents the time period when the user's operation is intensive, and the trough interval reflects the period when the operation is sparse. The identification of peak and trough is realized by sliding window statistical method. When the operation frequency in the window exceeds the preset density threshold, it is determined as peak interval, and when it is lower than the threshold, it is determined as trough interval.

[0061] The division of operation intensity level is based on the distribution of peak and trough intervals. High operation intensity level corresponds to user behavior in peak interval, indicating that the user is currently in an active state; low operation intensity level corresponds to low valley interval, reflecting that the user's interest is temporarily weakened or the attention is distracted. The dynamic change of operation intensity level is used to capture the interest fluctuation in user session. The extraction of depth feature is based on the hierarchical structure of the pages visited by the user in the current session, such as the jump path depth from the home page to the detail page, or the number of steps from search to the final click. The stay time feature is obtained by counting the user's stay time on a specific page. Longer stay time may reflect that the user has higher attention to the content.

[0062] 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 contains the timing information of user operations, but also covers the complexity of operation paths and the depth of content consumption. The dynamic updating mechanism of real-time behavior features ensures that newly generated user behavior data can be timely incorporated into the feature calculation process, avoiding interest prediction bias caused by data delay.

[0063] The generation of user behavior dynamic sequence relies on the construction of graph network model. The nodes in the graph network are defined as user operation types, such as click, play, purchase, etc., and the edges between nodes represent the transition relationship between operation types. The edge weight is calculated by counting the transition frequency between operation types in historical data. Operation pairs with high frequency transition have higher edge weight. The input of behavior delay features is realized by marking the time delay of operation transition, such as the time difference from clicking to purchasing or the interval length from browsing to collecting. 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 goods.

[0064] Graph convolution operation updates the node embedding representation, aggregates the information of adjacent nodes and preserves the timing characteristics of operation transition. The updated node embedding contains the context information of operation types, such as the tendency to browse related recommendations after clicking on a certain product. The operation path sequence with time label is generated by traversing the high weight edges in the graph network, and each node in the sequence is attached with timestamp information, forming the spatio-temporal expression of user behavior. The sliding window sampling technique processes the operation path sequence in segments, and the window size is dynamically adjusted according to the average duration of user session. The operation path segment in each window constitutes a subset of the user behavior dynamic sequence, which is used to analyze the interest change pattern in a short time.

[0065] The analysis of behavior dynamic sequence focuses on the stability and burstiness of operation transition. Stability is measured by repeated operation patterns in the sequence, such as frequent search-click-purchase path; burstiness is judged by new operation transitions, such as sudden increase in video playback behavior. The dynamic updating mechanism of sequence segment ensures that newly generated user behavior can be timely supplemented into the analysis process, avoiding incomplete interest capture due to data truncation. The output format of sequence segment adopts time-encoded vector form, which is convenient for subsequent interest fluctuation analysis algorithm processing.

[0066] The association analysis of operation intensity level and behavior dynamic sequence reveals the intensity variation law of user interest. The behavior dynamic sequence under high operation intensity level usually contains dense operation transitions and short delay features, reflecting the concentrated outbreak of user interest; the sequence under low operation intensity level presents scattered operation patterns and long delay features, indicating that the interest is in the flat or declining period. This association provides a reference basis for the calculation of interest fluctuation bias value in the intensity dimension.

[0067] The collaborative update mechanism of real-time behavior features and behavior dynamic sequences ensures their consistency in the time dimension. When new user behavior data is generated, real-time behavior features are updated first, and then the regeneration of behavior dynamic sequences is triggered. This collaborative mechanism avoids interest prediction errors caused by the time asynchronization of features and sequences, and improves the response speed of dynamic push strategies. The version control mechanism of features and sequences records the timestamp and change content of each update, facilitating the tracking of user interest evolution process.

[0068] The storage of user behavior dynamic sequences adopts a hierarchical structure, with raw operation data saved in the bottom database, sequence fragments stored in the intermediate cache layer, and aggregated sequence features placed in the high-level analysis module. This storage structure balances the data access speed and analysis depth requirements, ensuring that the real-time push system can quickly obtain the necessary sequence information. The index mechanism of sequence fragments organizes according to user session identification and time range, supporting efficient range queries and real-time update operations.

[0069] The dynamic adjustment of behavior delay features is based on statistical analysis of time difference in sequence fragments. When changes in the distribution of user operation intervals are detected, the threshold of delay features automatically adapts to the new time difference range. This adaptive mechanism avoids the invalidation of delay features caused by changes in user behavior patterns, such as the transition from rapid browsing to deep reading, and the system can recognize the new delay pattern and adjust the calculation method of interest lag accordingly. The incremental update algorithm of delay features only processes newly generated behavior data, reducing computational overhead while maintaining the timeliness of the features.

[0070] The anomaly detection module of operation path sequences identifies behavior fragments that do not conform to the regular pattern, such as repeated operations in a very short time or long periods of inactivity. The labeling and isolation of abnormal fragments prevent them from interfering with dynamic sequence analysis and trigger the reevaluation process of user attention. The threshold of anomaly detection is dynamically set according to the statistical distribution of user historical behavior, avoiding false positives or false negatives caused by fixed thresholds.

[0071] The precision control of time sequence labeling is achieved through high-resolution timestamps, ensuring the accurate order of events in the operation path sequence. The time synchronization protocol coordinates the time records of different data sources, eliminating the time sequence confusion caused by device clock differences. The standardized processing of time sequence labeling uniformly converts timestamps of different precision into a standard format, ensuring the consistency of the time dimension in sequence analysis.

[0072] The parameter optimization of the sliding window is dynamically adjusted by monitoring the coverage of the sequence analysis and the real-time requirement. A large window size will lead to a decrease in the sensitivity of interest changes, while a small window size may introduce noise interference. The automatic adjustment algorithm of the window parameter balances the integrity of the sequence fragments and the analysis granularity, and selects the optimal window size according to the stability degree of the user behavior. The step size of the window sliding controls the degree of overlap between the sequence fragments, and a moderate overlap ensures the continuity of the capture of interest changes.

[0073] The online learning mechanism of the graph network model continuously updates the weights of the nodes and edges to adapt to the changes in the user operation habits. The newly added operation type automatically expands the network nodes, and the decay function of the historical operation transition frequency reduces the influence of old data on the edge weights. The incremental training algorithm of the model only processes the newly added behavior data, avoiding the computational burden of retraining the full data. The dimension reduction technology of the network embedding retains the key operation transition patterns while reducing the computational complexity, improving the efficiency of dynamic sequence generation.

[0074] Embodiment 2: refer to Figure 3 The construction of the user preference prediction network starts from the branch input of static attribute features and real-time behavior features. The static attribute features include user demographic information, device attributes, historical preference labels, and other long-term stable features. These features are normalized to form a fixed-dimensional vector representation. The real-time behavior feature branch receives the dynamic behavior sequence features generated in Embodiment 1, including user's recent operation intensity, path depth, and stay duration, and other short-term behavior indicators. The input layers of the two branches use independent feature encoders to handle the dimension and distribution differences of different types of features.

[0075] The design of the feature cross layer calculates the correlation degree between static attributes and real-time behavior features using cosine similarity. The rows of the similarity matrix represent the static attribute feature dimensions, and the columns correspond to the real-time behavior feature dimensions. The matrix element values reflect the matching degree of the two types of features in a specific dimension. The feature vectors are normalized before similarity calculation to eliminate the influence of vector length on the similarity result. The high similarity area in the matrix indicates that there is a strong correlation between static attributes and current behavior, for example, users of a certain age group tend to produce a certain operation mode. The generation process of the similarity matrix uses batch calculation, which supports cross-analysis of multiple user features at the same time.

[0076] The feature interaction intensity coefficient is extracted from the similarity matrix and realized through matrix element aggregation. The calculation of the coefficient considers the importance difference of different feature dimensions, and gives higher aggregation weight to key dimensions. The dynamic adjustment mechanism of the interaction intensity coefficient monitors the stability of the feature cross result, and automatically recalibrates the coefficient calculation parameters when significant changes in the similarity distribution are detected. The output format of the coefficient is a multi-dimensional vector, and each dimension corresponds to the intensity quantization value of a specific feature cross combination.

[0077] The multilayer perceptron receives feature interaction intensity coefficients as input and learns the dynamic influence patterns 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 introduced into the generation process of the dynamically influencing weight vectors, highlighting the contribution of high-interaction-intensity features to the final weight allocation. Weight vector normalization ensures consistency in weight comparisons among different users, facilitating subsequent comprehensive preference calculations.

[0078] The distribution of interest tags in historical interaction records is obtained by statistically analyzing users' content category preferences over long-term behavior. The tag distribution data structure adopts a probability distribution form, with each tag corresponding to a preference intensity value, reflecting the user's relative interest in that type of content. The encoding process of interest feature vectors preserves the hierarchical relationship between tags; for example, the tags "sports-football" and "sports-basketball" are associated with the higher-level "sports" category. Sparse vector representation techniques filter out low-frequency noise tags, focusing on the user's main areas of interest.

[0079] Dimensionality reduction and compression employs a nonlinear manifold learning method to map high-dimensional interest feature vectors to a low-dimensional latent space. The 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. The update frequency of the implicit representation of low-dimensional interests is lower than that of real-time behavioral features, typically recalculated daily or weekly, balancing the need for long-term interest stability and adaptability. The storage of implicit representations uses an incremental update strategy, adjusting only some dimensional values ​​based on newly added interaction records, reducing computational overhead.

[0080] The concatenation of the dynamic influence weight vector and the low-dimensional interest implicit representation employs a gated fusion mechanism. The fusion process learns the complementary relationship between the two types of features, controlling the information flow in each dimension through a trainable parameter matrix. The generation of comprehensive preference features introduces feature importance ranking, prioritizing the retention of feature combinations most discriminative for the current scenario. Feature timeliness markers record the update time of each component, providing a freshness reference for subsequent push strategies.

[0081] The training process of the user preference prediction network employs a two-stage strategy. The first stage uses historical data to pre-train the network's basic structure, learning general patterns of feature interactions and weight mappings. The second stage uses online learning to fine-tune the network parameters based on real-time user feedback data. A weighted sampling method for training samples balances the coverage of different interest categories and user groups, preventing the model from favoring high-frequency behavior patterns. The network's version control mechanism retains historical parameter snapshots, allowing for rapid rollback to a stable version in the event of abrupt changes in interest distribution.

[0082] The monitoring system of the feature cross layer detects the distribution changes of the similarity matrix, identifies new associations between user behavior patterns and static attributes. When a significant new cross pattern is detected, the feature interaction intensity coefficient recalculation process is triggered. The monitoring indicators include the mean shift, variance change and singular value distribution of the matrix elements, and the stability of feature interaction is evaluated from multiple angles.

[0083] The adaptive learning rate adjustment of the multi-layer perceptron dynamically changes according to the update amplitude of the weight vector. When detecting rapid changes in user interest, the learning rate is automatically increased to speed up model adaptation; in the interest stable period, the learning rate is reduced to improve parameter convergence accuracy. The sparsity constraint of network parameters avoids overfitting problems, and the regularization term maintains the generalization ability of the model. The batch size of parameter update is dynamically adjusted according to system load, balancing calculation efficiency and model stability.

[0084] The time decay function of interest label distribution reduces the influence of old interaction records on the current label weight. The decay coefficient is differentiated according to the difference of content types, for example, the decay rate of news content is faster than that of video content. The cold start processing of labels assigns initial weights to newly emerging content categories, avoiding preference prediction blind spots caused by missing historical data. The dynamic expansion of label hierarchy supports seamless access to new content categories without retraining the entire label distribution model.

[0085] The dimension number of low-dimensional interest implicit representation is automatically adjusted according to user activity. High-active users' implicit representation uses higher dimensions to capture detailed interest differences, while low-active users use lower dimensions to prevent overfitting. The selection criteria for dimensions are based on the balance between reconstruction error and computational cost, and the optimal dimension configuration is determined through validation set evaluation. The visualization projection method of implicit representation helps understand the distribution of users in the interest space, assisting in the optimization decision of push strategy.

[0086] The version management of comprehensive preference features records the content changes of each update, supporting the tracing of interest evolution history. The feature difference comparison algorithm quantifies the distance between adjacent versions, and triggers the emergency adjustment process of the push strategy when a mutation update is detected. The distributed cache mechanism of features ensures fast response during high-frequency access, using the least recently used strategy to manage cache space.

[0087] The update audit mechanism of static attribute features ensures that long-term stable user information is not modified by short-term fluctuations. The verification process of attribute changes checks the compatibility of new data with historical records, and only when it is confirmed that the attribute has changed will the feature representation be updated. Special processing of sensitive attributes uses differential privacy technology to maintain the effectiveness of features while protecting user privacy.

[0088] Conflict detection between real-time behavior features and static attributes identifies contradictory situations, such as a user's real-time operation mode not matching their demographic attributes. Conflict resolution strategies select more reliable feature sources based on confidence evaluation results or initiate manual review processes. Analysis of conflict logs helps identify the root causes of system feature extraction or user behavior anomalies.

[0089] An interpretive analysis tool for the user preference prediction network visualizes feature cross paths and weight distribution logic. Analysis results display how static attributes influence final preference predictions through different levels in an interactive graphical form, enhancing model credibility and debugging efficiency. Extracting interpretive features identifies the most influential raw feature combinations on prediction results, assisting business understanding of key factors in user decision-making.

[0090] Elastic scaling of network structure automatically adjusts computing resource allocation based on request load. Horizontal expansion strategies increase parallel computing instances during peak periods, and reduce resources during idle times to lower costs. Dynamic scheduling algorithms predict workload trends at different times to prepare sufficient processing capacity in advance.

[0091] An automated feature engineering pipeline monitors changes in data distribution and triggers necessary feature recoding. Anomaly detection modules in the pipeline identify data quality issues during feature extraction, such as abnormal increases in missing values or sudden changes in numerical ranges. Automatic problem repair attempts to restore feature effectiveness using interpolation or rollback strategies, and serious cases trigger manual intervention alerts.

[0092] Confidence evaluation of preference prediction results quantifies the reliability of the current user prediction model. Low-confidence predictions trigger supplementary data collection or alternative prediction strategies to avoid unreliable results affecting the quality of the push. Confidence calculation considers factors such as feature integrity, model adaptability, and behavior novelty.

[0093] A closed-loop learning mechanism for user feedback uses actual interaction of pushed content as training samples back into the system. Weighted processing of feedback data distinguishes between explicit and implicit feedback with different information quantities. Negative sample mining strategies identify potential negative preferences of users who have not interacted with displayed content, balancing the proportion of positive and negative samples.

[0094] Multi-objective optimization handles trade-offs between different business objectives, such as the conflicting demands of click-through rate and dwell time. Configurable fusion parameters allow adjustment of optimization priorities according to different scenarios, meeting the requirements of differentiated push strategies. Dynamic priority of optimization objectives automatically adjusts according to changes in business indicators, maintaining the optimal balance of overall effectiveness.

[0095] Preference drift detection algorithms identify gradual trends in user interests. The drift adaptation mechanism progressively adjusts 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.

[0096] Cross-domain feature transfer learning leverages user data from other related domains to enhance preference predictions in the primary domain. The transfer process controls for negative transfer effects caused by domain differences, improving knowledge transfer efficiency through feature alignment and instance weighting. The selection criteria for the source domain are based on a comprehensive evaluation of its correlation strength with the target domain and its own data quality.

[0097] 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. Shorter segment lengths are used during high-density operation periods to capture subtle changes, while the segment lengths are appropriately extended during low-density periods to avoid data sparsity. The establishment of segment boundaries considers not only time intervals but also natural transition points between operation types, such as the time node when switching from browsing to purchasing behavior. The segmentation process preserves the temporal relationship of the original sequence, ensuring continuity between adjacent segments.

[0098] 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 statistically analyzed and normalized to a probability distribution. The entropy value is calculated using the following formula:

[0099]

[0100] in, Indicates time The entropy value, This represents the total number of operation types. It is the first The probability of a type of operation appearing in the current segment. Changes in entropy reveal the stability of user interests; low entropy indicates concentrated operation types, reflecting a clear interest focus; high entropy indicates scattered operations, with interests in an exploratory or wandering state. The rate of change of entropy is calculated by comparing the entropy differences between adjacent segments; positive values ​​indicate that interests tend to be scattered, while negative values ​​reflect that interests are gradually focusing.

[0101] The construction of the entropy value change rate sequence considers the time decay effect, and recent changes are given higher weights. The smoothing of the sequence eliminates the noise interference caused by short-term fluctuations and retains the trend changes. The introduction of the lag adjusts the time alignment of the entropy value change rate, compensating for the inherent delay between behavior occurrence and interest feedback. The adjusted change rate sequence is input into the time series prediction model, which uses the attention mechanism to enhance the ability to capture key turning points.

[0102] The structure of the time series prediction model includes an encoder and a decoder. The encoder converts the entropy value change rate sequence into a hidden state representation, capturing the long-term dependencies in the sequence. The decoder generates the interest fluctuation prediction curve for the future period based on the hidden state, with each point of the curve corresponding to the expected deviation value at a specific time. The model training uses a multi-step prediction strategy, forcing the learning of long-term dynamic patterns rather than simple one-step prediction. The confidence interval of the prediction curve is estimated by the Monte Carlo dropout technique, reflecting the uncertainty of the prediction results.

[0103] The preset value of the interest decay coefficient is derived from the long-term behavior statistics in the user profile, representing the natural decay rate of the interest of this type of user. The actual decay rate is calculated by fitting the entropy value change trend of the recent multiple analysis segments. The difference between the two generates the interest stability index, with a positive value indicating that the interest decays faster than expected, and a negative value reflecting that the interest persistence exceeds the average level. The calculation of the index uses a sliding window approach, with the window size being adaptively adjusted to match the user's behavior rhythm.

[0104] The generation process of the deviation correction parameter compares the matching degree of the prediction curve and the stability index. When systematic deviation is detected, the correction parameter adjusts the slope and amplitude of the prediction curve, making it more consistent with the actual observed interest change pattern. The correction process uses a gradual adjustment strategy to avoid prediction shocks caused by a single drastic correction. The corrected prediction curve is output as a continuous time function, supporting the interpolation calculation of the deviation value at any time.

[0105] The reevaluation mechanism of the interest analysis segment is triggered when an abnormal entropy value change is detected. The abnormality is determined based on the statistical characteristics of the historical segment distribution, with changes exceeding three standard deviations being considered abnormal. The reevaluation process temporarily excludes the influence of abnormal segments and re-includes them in the analysis after accumulating sufficient data. The credibility score of the segment guides the weight distribution of the subsequent prediction model input, and the influence of low credibility segments is appropriately suppressed.

[0106] Hierarchical clustering of operation types optimizes the granularity control of entropy value calculation. Similar operations are aggregated into higher-level categories to avoid distortion of entropy value fluctuations caused by too fine classification. The clustering criteria not only consider the surface similarity of operation names, but also analyze the transition probability and context relationship between operations. The dynamic clustering algorithm automatically expands the category system according to newly emerging operation types, maintaining the timeliness of the classification.

[0107] The online updating strategy of time series prediction model balances the contribution proportion of new and old data. Recent data obtains higher weight, but part of historical data is retained to maintain the generalization ability of the model. The updating frequency is automatically adjusted according to the prediction error, and the updating interval is shortened when the error increases. The version rollback mechanism of the model restores to the stable version when the performance decreases significantly, ensuring the continuity of the prediction service.

[0108] The normalization processing of entropy value change rate eliminates the absolute value difference between different users. The normalization benchmark adopts the moving average of the user's historical entropy value, highlighting the relative change rather than the absolute level. Cross-user comparison uses percentile conversion to map the original entropy value to a unified comparison scale. The periodic update of the normalization parameter adapts.

[0109] Embodiment 4: refer to Figure 5 The trigger condition of the real-time behavior feature recalibration module is that the user interest fluctuation deviation value exceeds the preset threshold. When the system detects that the deviation value of a certain user is higher than the threshold for three consecutive analysis periods, it starts the recalibration process. The process first separates the noise operation events from the original user behavior data set, such as a user usually browsing clothing products on an e-commerce platform, but suddenly there are multiple consecutive digital product quick click behaviors, and each click lasts less than 1 second. Such abnormal operations are marked as potential noise events, which are screened by the following rules: operation interval time is abnormally shorter than the historical average level, operation type and user portrait matching degree are too low, and operation sequence suddenly inserts incoherent behavior patterns.

[0110] The identification of noise events uses a multi-dimensional cross-validation method. The following table shows an analysis example of a user behavior data, which includes original operation records and noise judgment results:

[0111] Table 1: Analysis example of user behavior data

[0112] Timestamp Operation Type Page Category Stay Duration (seconds) Adjacent Operation Interval (seconds) Match Degree with Historical Preference Noise Label 2025-07-3110:05:31 Click Men's Clothing 12 8 0.92 Normal 2025-07-3110:05:43 Browse Men's Clothing 23 0 0.89 Normal 2025-07-3110:06:06 Click Mobile Phone 0.8 5 0.12 Noise 2025-07-3110:06:11 Click Notebook Computer 0.6 0 0.15 Noise 2025-07-3110:06:12 Return Home Page 2 1 0.45 Pending 2025-07-3110:06:14 Click Men's Clothing 15 2 0.91 Normal

[0113] The noise judgment in the table integrates multiple dimensions such as operation type, dwell time, interval time, and historical preference matching degree. The "pending" marked operation needs further analysis of its context relationship to determine the final classification. After the noise events are isolated, the system retains a complete copy of the original data for subsequent audit, and temporarily excludes the impact of these data in the processing process.

[0114] The data representation form update of user real-time behavior characteristics involves the re-allocation of feature weights. Taking a reading application as an example, the user originally focuses on technology articles, and the weight of the "technology" label in the feature vector is 0.85. Suddenly, there are a large number of rapid sliding behaviors of entertainment articles, which cause the weight of the "entertainment" label in the original feature vector to abnormally rise to 0.6. The re-calibration process reduces the contribution of such temporary abnormal behaviors, and the weight of the "entertainment" label in the adjusted feature vector returns to a reasonable range of about 0.2. The amplitude control of feature expression update adopts a gradual strategy, and each adjustment does not exceed 30% of the last version of the feature value, avoiding system shock caused by a single drastic change.

[0115] The re-computation of behavior delay characteristics focuses on the change of time pattern in normal operation sequence. For example, a user of a certain video platform, historical data shows that the average delay from browsing to collecting is 45 minutes, but recently there are a large number of operations that complete browsing to collecting within 3 minutes. After re-calibration, the system identifies that such rapid collection may be due to the improvement of content recommendation quality rather than the sudden change of user interest, so it divides the delay feature calculation into two modes: the regular browse-collect path keeps the original delay amount calculation, and the rapid collection path establishes a separate delay feature sub-model. This hierarchical processing more accurately reflects the interest feedback speed under different behavior patterns.

[0116] The optimization of dynamic influence weight is achieved by comparing the interaction intensity of static attributes and calibrated behaviors. Taking the age attribute as an example, the dynamic weight of "fashion" content for users aged 25-30 is originally 0.7, but a user's abnormal weight fluctuation to 0.3 due to temporary noise behavior. The re-calibration process restores the weight to a reasonable interval of 0.65 by analyzing the similarity of the user's behavior after calibration with other users of the same age group. The iterative process of weight optimization uses small step multiple updates, and checks the coordination with other related features after each iteration to ensure the rationality of the overall weight distribution.

[0117] The optimized user interest fluctuation bias value generation mechanism increases the robustness check link. The system compares the results of three calculation paths: based on complete original data, excluding all suspected noise, and partially excluding noise. When the difference between the bias values of the three paths exceeds the allowed range, the manual review process is triggered. Under normal circumstances, the optimized bias value takes the median of the three paths, balancing the strictness of noise filtering and the preservation of data integrity. The final output of the bias value is accompanied by a confidence score, reflecting the reliability of the current calculation result.

[0118] An example of the generation logic of content feature selection instructions: after re-calibration, a music user's interest fluctuation deviation value shows that the preference intensity for "electronic music" has decreased from 0.8 to 0.5, while the interest in "jazz" has increased from 0.3 to 0.6. The system generates instructions accordingly, requiring the content library to return content items with higher jazz feature weights than electronic music. The detailed parameters of the instructions include: primary tag weight range, secondary tag exclusion list, content freshness preference, etc. The priority of the instructions is set considering the change amplitude of the deviation value, and the interest points with large fluctuations obtain higher selection priority.

[0119] The content display interface of the terminal device processes the final presentation of the push queue. The interface optimizes the content layout according to the device type and screen size, with mobile phones prioritizing the display of concise information and tablets providing more comprehensive auxiliary content. The display logic takes into account the user's current operating environment, such as automatically reducing the priority of video content in a mobile network and restoring full recommendation weights in a WiFi environment. The feedback collection mechanism of the interface records the differences between actual user interaction and expected push, and these data will flow back to the evaluation link of the re-calibration module.

[0120] The execution process of dynamic update instructions uses a transaction mechanism to ensure data consistency. Update operations are either completely successful or completely rolled back to the previous stable state. The version control system records detailed logs of each update, including the trigger reason, modification content, and execution result. Changes to key parameters require double verification, especially adjustments to user core interest points must pass an additional consistency check.

[0121] The self-learning function of the noise separation module gradually refines the filtering rules. The system regularly analyzes the subsequent development of operations marked as noise to confirm whether they are truly abnormal behavior. Misjudgment cases will trigger parameter adjustments of the filtering rules, such as when certain types of short-term but real interest exploration behaviors are frequently misjudged as noise, the system automatically relaxes the judgment threshold for this type of operation. The speed control of the learning process avoids excessive adjustment of long-term accumulated experience rules due to short-term pattern changes.

[0122] The AB testing framework of the content push strategy verifies the adjustment effect. User traffic is randomly allocated to different strategy groups, and new strategies are only fully launched when they outperform the baseline. The test indicators not only include traditional indicators such as click-through rate, but also focus on comprehensive effects such as user long-term retention and satisfaction survey. The gray release mechanism of the strategy allows the coverage of new strategies to gradually expand, closely monitoring the trend of various indicators.

[0123] The abnormal recovery mechanism is automatically started when the system detects a continuous decline in push effectiveness. The recovery process first falls back to the last stable version of the push strategy, while in-depth analysis of the root cause of the problem. Common recovery operations include: resetting some feature weights, clearing outdated cache data, and strengthening noise filtering strength. User feedback during the recovery process is given higher weight, accelerating the system's adaptive adjustment speed.

[0124] The priority processing mechanism of user explicit feedback overrides the automatic push decision. When users actively collect, rate, or complain about certain content, these feedbacks immediately affect the adjustment of the current push queue without waiting for the regular recalibration period. The credibility assessment of explicit feedback distinguishes the weight of different feedback types, such as five-star ratings being more credible than simple clicks. Semantic analysis of feedback content extracts the user's direct intention, avoiding misinterpretation caused by simple statistics.

[0125] In embodiment 5, the generation process of the content push queue considers the terminal device's context environment parameters, including but not limited to device type, network status, geographic location, current time, battery level, and ambient light. Device type information determines the layout and interaction mode of content display, such as single-column waterfall display for smartphones and grid layout for smart TVs. Network status parameters are divided into three basic modes: mobile network, WiFi, and offline. The content loading strategy differs significantly under different states. In mobile network environment, text and low-resolution image content are preferentially pushed, while in WiFi environment, the default weight of high-definition images and videos is restored. Geographic location information is used to adjust the priority of region-related content, such as the display weight of local news and life service, which dynamically changes according to the user's location.

[0126] Context environment parameters are collected in real time through device API, with sampling frequency dynamically adjusted according to parameter type. Network status and power parameters that change quickly are sampled at high frequency, while device type and screen size parameters that are stable are only collected once during initialization. Normalization of parameters eliminates the influence of different dimensions, mapping various parameters to a unified scoring scale. Time 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 effects, such as low battery and weak network signal, which require special optimization of content strategy.

[0127] The calculation of the display weight coefficient is based on the matching degree of the context environment parameters and the content features. Each content vector contains multiple display-related features, such as text length, picture proportion, video duration, 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 long text content on a smart watch will be appropriately reduced. Secondly, the influence of network status on content loading speed is considered, and content with high bandwidth demand is downgraded in mobile network environment. Finally, the ambient light and time information are considered to adjust the display form, and the contrast and brightness parameters of the content in night mode are automatically optimized.

[0128] The extraction process of the inherent feature value of the content vector considers the attributes and historical performance of the content itself. Text content analyzes keyword density and sentiment orientation, picture content identifies main visual elements and color distribution, and video content extracts key frame features and rhythm changes. Historical performance data includes average click rate, completion rate, and sharing times of similar content. The dynamic updating mechanism of the inherent feature value tracks the changing trend of the content life cycle, and the initial value of newly released content is set based on the average performance of similar content, and gradually adjusted to the actual data as the exposure increases. The smoothing processing of the feature value avoids the drastic change of the display weight caused by short-term fluctuations.

[0129] The generation process of the push content instance with time label integrates the display weight coefficient and the inherent feature value. The time label reflects the best display time period of the content, such as the explicit start and end time of the limited-time promotion content, and the golden time period of the recommended reading of news information. The generation of the instance uses a templating method, and applies a preset display template according to different content types, while leaving a certain degree of adaptive adjustment space. Template parameters include title font size, picture cropping ratio, and interactive button position details, which are automatically optimized according to the device environment. The verification process of the instance checks whether the generated result meets the display specification, avoiding layout disorder or content truncation, etc.

[0130] The insertion position of the push content instance in the queue is determined by the time label and the comprehensive priority. Content with strict time limit gets the privilege of inserting in front of the queue, and instances close to the deadline are automatically promoted in the sorting position. The sorting of regular content considers multiple factors such as user interest matching degree, content freshness, and diversity balance. The real-time adjustment mechanism of the queue responds to changes in environmental parameters, for example, when it is detected that the network is switched from mobile data to WiFi, the position of video content is automatically moved forward. The conflict resolution of the insertion position adopts a hierarchical strategy, and the same level content is sorted by time, and different level content allows high priority instances to occupy the position.

[0131] The content display interface of the terminal device handles the final rendering of the queue instance. The interface optimizes the resource loading order according to device performance, with first-screen content loaded first and non-visual area content using lazy loading technology. The adaptive adjustment of the rendering process considers the actual display effect, and when it is detected that certain content elements cannot be normally displayed, a degraded display solution is automatically enabled. The interaction tracking module of the interface records user behavior data such as scroll depth, dwell time, and operation path, which are fed back to the queue generation link for optimization of subsequent pushing.

[0132] The time stamp conflict resolution process handles the case where the time requirements of multiple content instances overlap. The conflict detection algorithm identifies the overlapping part of the time window and determines the priority according to business rules. For example, the time conflict of promotional content usually adopts the first-come-first-served principle, while urgent notification type content can interrupt the display of regular content. Conflict resolution records detailed logs for subsequent analysis and rule optimization reference.

[0133] Adaptive adjustment of device performance optimizes the running performance of content on different hardware. Low-performance devices receive simplified content instances, with unnecessary animation effects and background loading turned off. The performance monitoring module tracks the content rendering frame rate and response delay, and automatically triggers the optimization process when performance bottlenecks are detected. Long-term statistics of performance parameters establish a device capability profile, providing a basis for subsequent prediction.

[0134] The display coordination of multi-modal content handles the presentation order of mixed type content. For example, the alternating display of text and image mixed content and pure video content needs to consider the user's reading rhythm and attention shift. The coordination algorithm analyzes the consumption patterns of different types of content in the user's historical behavior, and finds the optimal display interval and transition method. The associated pushing of multi-modal content appropriately combines content with different forms but related semantics, providing a more complete information acquisition experience.

[0135] Content display optimization considering the influence of circadian rhythm on user preference. Daytime mode uses high-contrast bright colors, and nighttime mode switches to low-blue light dark theme. The data fusion of light sensors integrates time information, avoiding unnecessary theme switching caused by short-term environmental changes. The de-bouncing processing of the adaptive algorithm eliminates the flicker problem of the interface caused by rapid light fluctuations.

[0136] Battery state-aware content strategy adjustment prolongs device usage time. In low battery state, reduce background data prefetching and content update frequency, and suspend unnecessary visual effects. The power prediction model estimates the remaining available time based on the current usage mode, and adjusts the aggressiveness of content pushing accordingly. The compensation mechanism after power recovery gradually restores full functionality, avoiding sudden surge of resource demand.

[0137] Geolocation related service content push considers user movement trajectory. Local content in the resident area remains stable display, while temporary location in the travel route pushes destination related information. Trajectory prediction algorithm analyzes user's moving speed and direction, and predicts the area service that may be interested. Geofencing technology triggers content switching in specific area, for example, automatically promoting the priority of merchant preferential information after entering the mall.

[0138] Cross-device content synchronization maintains the continuity of user experience. User browses content halfway on mobile phone, and can continue from the same position on tablet device. Synchronization mechanism identifies the breakpoint position of content, records user's reading progress and operation state. Conflict resolution strategy usually takes the last updated device data as the standard when multiple devices modify content state at the same time, and key operations need explicit confirmation.

[0139] Personalization recognition of multi-account devices correctly handles the scenario of family shared devices. User recognition technology distinguishes different users on the same device, and switches corresponding content push strategy according to login state or behavior characteristics. Limited push in guest mode provides general content, protecting the privacy preferences of the main user. Account association intelligent recognition establishes content recommendation association of family members under user permission, and discovers possible common interest points.

[0140] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action, without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0141] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the spirit and scope of the application in its broadest form. The scope of the application is defined by the following claims and their equivalents.

Claims

1. A method for real-time personalized content push based on user profiles, characterized in that, Includes the following steps: Collect user behavior datasets and historical interaction records; extract real-time user behavior features from the user behavior datasets; and generate dynamic user behavior sequences based on the real-time user behavior features. Based on the changing trend of the dynamic sequence of user behavior, multiple behavioral delay features are identified, and the behavioral delay features are converted into the lag of user interest feedback. Obtain static attribute features from the user profile, and determine the dynamic influence weight of attribute features on behavioral features by combining the interaction relationship between the static attribute features and the real-time behavioral features. The user interest tendency is predicted in real time by using the lag 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, a dynamic update instruction for the user profile is triggered and the content push strategy is adjusted. The determination of the dynamic influence weights of attribute features on behavioral features includes: Construct a user preference prediction network, with the input layer containing static attribute feature branches and real-time behavior feature branches; A feature cross-layer is set in the user preference prediction network to calculate the cosine similarity matrix between static attribute features and real-time behavior features; Generate feature interaction strength coefficients based on the cosine similarity matrix; The feature interaction intensity coefficients are mapped into dynamic influence weight vectors using a multilayer perceptron. The generated user interest fluctuation deviation value includes: The dynamic sequence of user behavior is processed by time segmentation to divide it into multiple interest analysis segments; the rate of change of entropy value of operation type distribution is calculated within each interest analysis segment; Extract the entropy change rate sequence of the multiple interest analysis segments; The entropy change rate sequence and the hysteresis quantity are input into the time series prediction model; The time series prediction model outputs a continuous prediction curve of the user interest fluctuation deviation value. The time series prediction model performs: Obtain the preset value of the interest decay coefficient from the user profile; Calculate the actual decay rate of the operation type distribution in the interest analysis segment; An interest stability index is generated based on the difference between the actual decay rate and the preset value of the interest decay coefficient. The interest stability index is compared with the user interest fluctuation deviation value prediction curve, and the deviation correction parameter is output.

2. The method for real-time personalized content push based on user profiles as described in claim 1, characterized in that, The step of extracting real-time user behavior features from the user behavior dataset includes: Calculate the time interval sequence of continuous user actions; Identify the peak and trough intervals in the time interval sequence; classify the user's operation intensity level based on the peak and trough intervals; Extract the depth features and dwell time features of the current session based on the user's operation intensity level; The depth feature and dwell time feature are fused into the user's real-time behavior feature.

3. The method for real-time personalized content push based on user profiles as described in claim 1, characterized in that, The step of generating a dynamic sequence of user behavior based on the real-time user behavior characteristics includes: Establish a user behavior graph network model, where nodes represent user operation types and edge weights represent the transition frequency between operation types; The multiple behavioral delay features are input into the user behavior graph network model; Update the node embedding representation through graph convolution operations to generate a sequence of operation paths with temporal tags; The operation path sequence is sampled using a sliding window to output a dynamic sequence fragment of user behavior.

4. The method for real-time personalized content push based on user profiles as described in claim 1, characterized in that, The user preference prediction network also performs: Extract the distribution of interest tags from users' historical interaction records; The interest label distribution is encoded into an interest feature vector; The interest feature vector is subjected to dimensionality reduction and compression to generate a low-dimensional implicit interest representation; The low-dimensional implicit interest representation is weighted and concatenated with the dynamic influence weight vector to form a comprehensive preference feature.

5. The method for real-time personalized content push based on user profiles as described in claim 1, characterized in that, The condition when the user interest fluctuation deviation value exceeds the preset fluctuation threshold further includes: Start the real-time behavior feature recalibration module to separate noisy operation events from the original user behavior dataset; Update the data representation of the user's real-time behavior characteristics; recalculate the behavior delay characteristics and dynamic impact weights; An optimized user interest fluctuation deviation value is generated by using updated behavioral delay features and dynamic influence weights.

6. The method for real-time personalized content push based on user profiles as described in claim 1, characterized in that, The adjusted content push strategy includes: Content feature selection instructions are generated based on the optimized user interest fluctuation deviation value; Filter the set of content vectors that match the feature selection instructions from the content feature library; The content vector set is prioritized and sorted to generate a content push queue; the content push queue is then sent to the content display interface of the terminal device.

7. The method for real-time personalized content push based on user profiles as described in claim 6, characterized in that, The generated content push queue includes: Extract the context parameters of the user terminal device; Adjust the display weight coefficient of the content vector according to the aforementioned contextual parameters; By combining the display weight coefficient with the inherent feature value of the content vector, a push content instance with a time-sensitive tag is generated; Insert the push content instance into the specified position of the content push queue according to the time expiration mark.

Images