Potential keyword mining prediction method based on time sequence search behavior analysis

Abstract

The present application relates to a potential keyword mining prediction method based on time sequence search behavior analysis, aiming to solve the problem that the existing behavior mining system is not sensitive to interest change recognition and the keyword score is not adjusted in time. Its core technical scheme includes: slicing the search behavior data based on the sliding window, generating a multi-dimensional activity vector, and quantitatively representing the interest evolution uncertainty by using information entropy; constructing a short time sequence entropy sequence, generating a personalized entropy threshold by dynamic boundary constraint, and determining the interest disturbance event in real time; combining the keyword co-occurrence graph and the user jump path, and correcting the keyword score by a lightweight neural network submodule; collecting the update log feedback to realize the iterative optimization of the threshold generation logic. Compared with the prior art, the present application can improve the accuracy and response speed of interest pattern disturbance detection, realize adaptive adjustment of keyword score, and guarantee the real-time performance and effect optimization closed loop of the model.

Description

Technical Field

[0001] This invention relates to the field of time-series user behavior modeling and keyword mining and prediction technology, and in particular to a method for mining and predicting potential keywords based on time-series search behavior analysis. Background Technology

[0002] Currently, potential keyword mining and scoring prediction systems based on time-series user behavior modeling are gradually becoming an important technical support for improving internet search and recommendation services in both academia and industry. Most existing mainstream technical solutions employ static periodic scheduling mechanisms or fixed threshold triggering methods to periodically refresh or recalculate parameters of the keyword scoring model. Representative architectures include those based on statistical rule extraction, periodic time-series clustering, tag frequency threshold filtering, and some batch incremental update methods based on shallow time-series differencing. These solutions can effectively achieve coarse-grained dynamic maintenance of the keyword database and improve the system's initial adaptability to large-scale user behavior flows, but they face many limitations in practical applications.

[0003] Currently, the industry generally relies on fixed-period strategies or static threshold triggering mechanisms to refresh keyword scoring models. For example, some internet search platforms automatically rebuild their scoring models every 24 hours or week, or trigger large-scale recalculations based on pre-set behavioral indicator thresholds (such as keyword click-through rate, search volume, bounce rate, etc.). The advantage of this model is that it automates business processes and facilitates management and monitoring. However, at the same time, fixed periods and static thresholds cannot reflect changes in user behavior and interest migration characteristics on a fine-grained time scale. Therefore, frequent full refreshes can easily lead to wasted server resources and computational strain, while excessively long periods may result in keyword scoring lag, failing to respond promptly to sudden changes in user behavior and the emergence of new words.

[0004] In academia, while typical temporal behavior modeling methods (such as event flow clustering, temporal anomaly detection, and simple trend prediction) can monitor some behavioral traffic anomalies and interest shifts, their scoring update mechanisms mostly still rely on periodic tasks, fixed-point threshold judgments, or static discrimination logic based on the distance between cluster centers. This limits the model's ability to dynamically perceive user's individual behavioral characteristics and the activity level within a window, as well as its adaptive response capabilities. Although some systems have attempted to introduce lightweight incremental learning to fine-tune parameters for local periods of high behavioral activity, the entire scoring update process is often still constrained by the main cycle scheduling framework, lacking self-regulating capabilities driven by real-time temporal statistical signals.

[0005] Existing technologies are applied in areas such as maintaining internet search keyword databases, mining interest tags, predicting hot word trends, and updating user profiles for recommendation systems. In these scenarios, traditional scoring update mechanisms can meet the business needs of full keyword refreshes or periodic scoring calibration, providing coarse-grained management of large to medium-sized user behavior flows. However, for user groups with frequent behavioral changes and extremely rapid interest shifts, existing mechanisms often fail to capture local behavioral disturbances in a timely manner. This results in long-term lag in keyword scoring, insufficient response to less popular keywords, unoptimized resource consumption, and scoring changes after the emergence of new keywords failing to accurately reflect actual trends.

[0006] Technical analysis reveals the following shortcomings in existing user behavior modeling and keyword scoring update mechanisms: The scoring update process is strictly constrained by a periodic task framework, lacking adaptive refresh capabilities based on user behavior activity, and unable to determine the timing and scope of model recalculation in real time according to the entropy change trend within a single window. Full model reconstruction consumes significant resources, and insufficient response to local perturbations can lead to delays in new word cold starts or insensitivity to interest migration. Static threshold triggering and periodic refreshes cannot reflect the discrete evolution of user behavior and interest fluctuations within fine-grained time windows, resulting in delayed adjustments to high-value keyword scores, long-term neglect of niche keywords, and difficulty in dynamically activating long-tail keyword trends. Some incremental methods rely on simplified rules or static clustering results when correcting local parameters, lacking dynamic perception of the statistical uncertainty of temporal activity and mechanisms for fine-tuning local features.

[0007] In summary, current time-series search behavior analysis and keyword scoring prediction systems urgently need a technical solution that can use the uncertainty of behavioral activity (such as changes in local information entropy) as the core decision signal to achieve adaptive local updates of the scoring model, in order to overcome the shortcomings of traditional solutions such as lag, resource waste, and insufficient response. The industry urgently needs a novel adaptive dynamic scoring update mechanism that is highly accurate, consumes few resources, and is sensitive to sudden changes in local behavior. This patent aims to solve the key technical problem of the lack of dynamic adaptability in scoring model update mechanisms, providing the industry with a more cutting-edge, resource-sensitive, and interest-refined response solution. Summary of the Invention

[0008] This application provides a method for mining and predicting potential keywords based on time-series search behavior analysis, aiming to solve one of the problems or issues of the existing technology mentioned in the background section above.

[0009] The potential keyword mining and prediction method based on time-series search behavior analysis provided in this application specifically includes: S1: Obtain user search behavior stream data, and slice the search behavior stream data based on the sliding time window mechanism to generate a four-dimensional indicator behavior activity vector; S2: Use the information entropy calculation formula to perform discrete quantization on the behavior activity vector to obtain the single-window behavior activity entropy value; S3: Construct a short time series from multiple consecutive single-window behavior activity entropy values ​​and input it into the time series difference module to perform dynamic boundary constraint operation and generate a personalized entropy threshold that drifts with the rhythm of individual user behavior. S4: Determine whether the single-window behavior activity entropy value of the current window exceeds the corresponding personalized entropy threshold. If the determination result is that it exceeds the threshold, trigger the behavior pattern disturbance event identification process and perform influence domain analysis to obtain the set of keywords to be updated. S5: Construct a lightweight trend modeling submodule consisting of gated linear units superimposed with a residual attention mechanism, and take the set of keywords to be updated as input to perform local gradient correction; S6: Based on the linear constraint of the magnitude of the single-window behavior activity entropy value, the lightweight trend modeling submodule performs local gradient correction operation only on the set of keywords to be updated to generate keyword scoring results, and completes the prediction of potential keywords based on the keyword scoring results. S7: Record update log data and inject the update log data as feedback into the data source; S8: Perform iterative optimization processing on the generation logic of the personalized entropy threshold based on the feedback injection data source.

[0010] The potential keyword mining and prediction method based on time-series search behavior analysis provided in this application has the following beneficial effects: (1) By constructing an adaptive rating update mechanism driven by the inherent evolutionary uncertainty of user behavior, this scheme significantly overcomes the technical defects of traditional recommendation systems, which suffer from both response lag and over-update due to reliance on fixed-period scheduling or static threshold judgment. In existing technologies, model parameter updates are usually triggered based on preset time intervals or uniform thresholds, which are difficult to adapt to the heterogeneity of different user behavior rhythms. Especially when facing dynamic scenarios such as sudden interest migration or short-term high-frequency exploration, update delays or redundant calculations are very likely to occur. This scheme, on the other hand, achieves fine-grained perception of user behavior pattern perturbations by jointly encoding multi-dimensional behavioral features (including keyword frequency, behavioral type diversity, session interval distribution, and cross-device consistency) within the sliding time window and introducing information entropy to quantify the degree of activity dispersion. Furthermore, a lightweight temporal difference module is used to model entropy change trends, generating dynamic thresholds constrained by the individual's historical fluctuation envelope, current growth rate, and group benchmark. This allows the trigger boundary to adaptively drift with the rhythm of user behavior, thereby significantly improving sensitivity while maintaining a low false trigger rate. This ensures that the update process is initiated only when a semantic-level behavioral transition actually occurs, effectively improving the accuracy and timeliness of the model response.

[0011] (2) To address the high computational overhead and system load issues caused by traditional full-update strategies, this solution innovatively proposes a three-level linkage mechanism of "entropy-sensitive triggering - local influence domain identification - incremental scoring calibration," which significantly reduces resource consumption while ensuring update effectiveness. When a behavioral pattern disturbance event is detected, the system does not retrain all keywords. Instead, based on the keyword co-occurrence graph and recent click path analysis, it quickly locates the semantic subset most significantly affected by the disturbance, forming a set of keywords to be updated, thus achieving spatial focusing of the influence range. Subsequently, the lightweight trend modeling submodule is called, using gated linear units combined with residual attention structures. Without reconstructing the backbone network, only key parameters such as time decay coefficient, interest migration weight, and context stability factor are locally gradient corrected, and the correction magnitude is linearly adjusted by the current entropy value. This makes the adjustment more aggressive in high uncertainty scenarios and more conservative in low disturbance scenarios, balancing flexibility and stability. This design avoids the repetitive inference overhead of complex deep architectures (such as LSTM and Transformer), has good online deployment capabilities, and is suitable for efficient iteration requirements in large-scale real-time recommendation scenarios.

[0012] The aforementioned mechanisms collectively construct a closed-loop evolutionary system of "perception-decision-execution-feedback," which not only achieves a fundamental shift from passive, timed updates to proactive, on-demand responses, but also continuously optimizes future triggering strategies by feeding back the effect data of each update (such as keyword conversion improvement rates) to the dynamic threshold generation module, thereby enhancing the system's self-evolutionary capabilities. The entire process eliminates the need for manually setting rule templates, static clustering, or explicit periodic configurations, freeing it from reliance on high-cost architectures such as collaborative filtering, graph neural networks, and BERT semantic modeling. This forms a new update paradigm that achieves high precision, low latency, and strong robustness under resource-constrained conditions. It is particularly suitable for application scenarios requiring rapid capture of user intent shifts, such as personalized search and efficient, diverse, and accurate matching of on-campus and off-campus resources in campus football event information recommendations, effectively improving the relevance, diversity, and consistency of recommendation results with user experience. Attached Figure Description

[0013] Figure 1 This is the main flowchart of a potential keyword mining and prediction method based on time-series search behavior analysis.

[0014] Figure 2 This is a sub-flowchart of a potential keyword mining and prediction method based on time-series search behavior analysis.

[0015] Figure 3 This is another sub-flowchart of the potential keyword mining and prediction method based on time-series search behavior analysis. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0018] like Figure 1 As shown, this application provides a method for mining and predicting potential keywords based on time-series search behavior analysis, specifically including: S1: Obtain user search behavior stream data, and slice the search behavior stream data based on the sliding time window mechanism to generate a four-dimensional indicator behavior activity vector; S2: Use the information entropy calculation formula to perform discrete quantization on the behavior activity vector to obtain the single-window behavior activity entropy value; S3: Construct a short time series from multiple consecutive single-window behavior activity entropy values ​​and input it into the time series difference module to perform dynamic boundary constraint operation and generate a personalized entropy threshold that drifts with the rhythm of individual user behavior. S4: Determine whether the single-window behavior activity entropy value of the current window exceeds the corresponding personalized entropy threshold. If the determination result is that it exceeds the threshold, trigger the behavior pattern disturbance event identification process and perform influence domain analysis to obtain the set of keywords to be updated. S5: Construct a lightweight trend modeling submodule consisting of gated linear units superimposed with a residual attention mechanism, and take the set of keywords to be updated as input to perform local gradient correction; S6: Based on the linear constraint of the magnitude of the single-window behavior activity entropy value, the lightweight trend modeling submodule performs local gradient correction operation only on the set of keywords to be updated to generate keyword scoring results, and completes the prediction of potential keywords based on the keyword scoring results. S7: Record update log data and inject the update log data as feedback into the data source; S8: Perform iterative optimization processing on the generation logic of the personalized entropy threshold based on the feedback injection data source.

[0019] Step S1: Acquire user search behavior stream data, and slice the search behavior stream data based on a sliding time window mechanism to generate a four-dimensional behavior activity vector. Specifically, this includes: S1.1: Real-time collection and cleaning of raw user search behavior stream data to remove invalid session records and standardize timestamp format, generating a structured behavior event sequence containing keyword identifiers, behavior type labels, session start time, and terminal device identifiers.

[0020] The received raw user search behavior stream data is collected and cleaned in real time to ensure that the input data can fully cover the user's search behavior events and maintain the accuracy of the time sequence.

[0021] Call the collection interface for multi-source heterogeneous data, perform parallel crawling operations on desktop browser logs, mobile application tracking data and cross-platform search API return results, and generate raw behavioral event streams that are initially sorted by event occurrence time.

[0022] The invalid session detection submodule is invoked on the collected raw behavior event stream. Based on the judgment rules of session duration being lower than the threshold, keywords being empty strings, behavior type labels being missing, or terminal device identification being abnormal, invalid record removal processing is performed to form a valid behavior event stream.

[0023] For all timestamp fields in the valid action event stream, the time format standardization method is called to uniformly convert them to Coordinated Universal Time (UTC) format and add millisecond-level precision to ensure the consistency of time boundaries for subsequent window slices.

[0024] The keyword text contained in the effective behavior event stream is mapped to a unique keyword identifier using field mapping operations, and the interactive behavior description is mapped to a standardized behavior type label. At the same time, the session start time field and the terminal device identifier field are bound according to a predetermined data structure to form a structured behavior event entity.

[0025] Through the above collection and cleaning methods, the original disordered search behavior flow is transformed into a structured sequence of behavior events containing keyword identifiers, behavior type tags, session start time, and terminal device identifiers, thereby ensuring data reliability and standardized feature output.

[0026] S1.2: Based on the preset time span parameter, perform a sliding time window slicing operation on the structured behavior event sequence to divide the continuous behavior flow into multiple time segments with overlapping regions, and generate a discretized behavior window dataset with independent time boundary constraints.

[0027] Receive structured behavioral event sequence data generated from step S1.1, including basic field information such as keyword identifier, behavior type label, session start time, and terminal device identifier.

[0028] A sliding time window slicing control mechanism is constructed based on preset time span parameters. A continuous sequence of structured behavioral events is used as the input object, and two core parameters, window length and window stride, are set to ensure that there is a controllable overlapping area between windows to improve the accuracy of time series analysis.

[0029] The window boundary determination operation is performed using the timestamp field. The event segment to which it belongs is determined by comparing the window start time and end time. The event is then assigned to the event set of the corresponding window one by one using a position index mapping table.

[0030] The window overlap rule module is invoked to perform cross-checking on the time boundaries of adjacent windows, ensuring that events within the overlap interval can be assigned to multiple windows simultaneously, thereby preserving the continuity of user behavior across boundaries in a short period of time.

[0031] For each event set within a window, an independent time boundary constraint record is generated, recording the absolute timestamp information of the window's start and end, and encapsulating it into a discretized behavior window data object to ensure that subsequent statistical processing steps are executed independently within the window's time range.

[0032] By using a sliding time window slicing method, the structured behavioral event sequence output in step S1.1 is transformed into multiple discrete behavioral window datasets with overlapping regions and independent time boundary constraints. This enables refined time-series slicing of the user search behavior flow, providing an accurate time-slicing basis for subsequent indicator statistics and feature quantification.

[0033] S1.3: The frequency of keyword occurrence in each discrete behavior window dataset is counted using statistical aggregation methods, and the distribution uniformity of behavior type labels is quantified by combining the Shannon diversity index calculation method, generating a basic statistical feature set representing content popularity and interaction richness.

[0034] For the discretized behavioral window dataset after slicing, read all keyword identifiers within the window and call the statistical aggregation method to perform frequency accumulation calculation to obtain the window-level keyword frequency vector.

[0035] The frequency vector of the keywords is indexed and mapped sequentially to match the corresponding set of behavior type labels. A diversity distribution table of behavior type labels within the window is constructed, and the occurrence frequency of each type of label is counted to form a category frequency sequence.

[0036] The frequency sequences of the categories are summed to obtain the total number of behavioral events within the window. Based on the ratio of the total number of behavioral events to the category frequency data, the proportion vector of each behavioral type is calculated and used as the input probability distribution of the Shannon diversity index.

[0037] The Shannon diversity calculation formula is used to perform multidimensional entropy aggregation on the proportion vector. The formula is as follows:

[0038] in, For the first Percentage of different behavior types This represents the diversity entropy value.

[0039] The frequency vector of the keywords and the diversity entropy value are combined into a two-dimensional feature group, and a matrix splicing method is used to form a basic statistical feature group that represents the popularity and richness of window content.

[0040] This processing method transforms the slicing results from the previous step into quantifiable keyword popularity parameters and behavioral interaction diversity indicators, enabling a preliminary characterization of the evolution of user interests.

[0041] S1.4: Calculate the session interval distribution variance based on the time difference between adjacent events in the discretized behavior window dataset, and simultaneously calculate the cross-device consistency ratio through device identifier matching logic. Then, fuse and map the basic statistical feature group, session interval distribution variance, and cross-device consistency ratio into an initial multidimensional index set.

[0042] Using the time difference between adjacent events in the discretized behavior window dataset as input variables, a session interval sequence is obtained by performing a difference operation on the timestamps of each pair of adjacent events. The variance calculation formula is applied to the session interval sequence, and numerical squared difference accumulation and normalization are used to obtain the session interval distribution variance, reflecting the degree of dispersion of the time distribution. Matching and comparison are performed on the terminal device identifier field extracted from the discretized behavior window dataset to establish a cross-device access mapping table, used to statistically analyze the behavioral consistency ratio of the same user across different devices. The cross-device behavioral consistency ratio is calculated using the ratio of the total number of matched sessions to the total number of sessions to obtain a numerical index, which serves as a cross-device consistency feature. The basic statistical feature groups generated in S1.3, such as keyword frequency and behavioral type diversity, are subjected to feature fusion mapping processing with the aforementioned session interval distribution variance and cross-device consistency ratio. Through the construction of a multi-dimensional feature fusion matrix, the three heterogeneous feature sources are mapped to the same coordinate system, resulting in an initial multi-dimensional index set containing four types of information: content popularity, interaction richness, temporal stability, and device consistency. By using feature fusion processing, the basic statistical feature set from the previous step is transformed into a multi-dimensional data set that can simultaneously reflect the temporal distribution characteristics of user behavior and the consistency status across devices, thereby achieving a more comprehensive characterization of the evolution of user interests.

[0043] S1.5: Perform normalization and joint encoding operations on the initial multidimensional index set to eliminate dimensional differences and compress the four heterogeneous indexes into a fixed-dimensional numerical array to generate a behavioral activity vector that characterizes the evolution of user interests within a single window.

[0044] Step S2: The behavior activity vector is discretely quantized using the information entropy calculation formula to obtain the single-window behavior activity entropy value. Specifically, this includes: S2.1: Obtain the behavior activity vector generated by the previous steps, which includes four-dimensional indicators: keyword frequency, behavior type diversity, session interval distribution, and cross-device consistency. Perform normalization mapping on the original values ​​of each dimension in the behavior activity vector to eliminate dimensional differences and generate standardized behavior probability distribution base data.

[0045] S2.2: Based on the standardized behavioral probability distribution basis data, a negative logarithmic weighting operation is performed on each dimension component using a logarithmic transformation method to map the linear probability space to the information metric space and generate an intermediate information sequence containing the independent information contribution of each dimension.

[0046] A dimensional iterative access mechanism is established for the standardized behavioral probability distribution basis data to read the four-dimensional components one by one and ensure that the data format meets the input requirements of subsequent information measurement space mapping.

[0047] The logarithmic transformation method module is invoked to perform natural logarithmic operations on each component dimension, and the result is transformed into a negative logarithmic value based on the negative sign control logic, thereby realizing a single-point mapping from the probability space to the information measurement space.

[0048] By using a weighted processing unit, the negative logarithmic values ​​of each dimension are multiplied with the corresponding original probability components to ensure that both probability weights and information intensity signals are preserved in the information content calculation.

[0049] A unified format mapping generator is used to collect the product results of each dimension as intermediate information components, and construct an intermediate information sequence containing four elements according to the dimension index order.

[0050] The intermediate information sequence is encapsulated into a structured information sequence object that meets the input requirements of the subsequent entropy aggregation mechanism through the data structure encapsulation module, thereby realizing the transformation from multidimensional heterogeneous vectors to a unified information representation.

[0051] Through the above chain-like derivation process, the standardized probability distribution basis data from the previous step is transformed into an intermediate information sequence containing independent information contributions from each dimension, thus achieving the expected technical effect of mapping the linear probability space to the information measurement space.

[0052] S2.3: The intermediate information sequence is weighted and summed using the Shannon entropy aggregation mechanism to comprehensively reflect the overall dispersion of the four-dimensional indicators within a single time window and generate an initial entropy parameter that characterizes the disorder of user behavior.

[0053] The processing object of the intermediate information sequence is initialized to ensure that it contains a four-dimensional independent information contribution component generated by the weighted operation of the preceding negative logarithm, and maintains the same dimension as the behavior activity vector.

[0054] The core operator of the Shannon entropy aggregation mechanism is invoked, and a weighting coefficient corresponding to the index weight is set for each dimension component. Based on the weight mapping table, it is ensured that the contribution ratio of different indicators in the entropy calculation conforms to the preset analysis model.

[0055] The weighted product operation is performed, multiplying the contribution of each dimension of information by the corresponding weight coefficient to form a set of product results, which serves as the direct input data for subsequent summation and aggregation.

[0056] The aggregation summation operator is used to sum the product result set to form a weighted total information value that characterizes the overall dispersion of the four-dimensional heterogeneous index.

[0057] The weighted total information value is compared with the standardization coefficient to ensure that the value range of the entropy parameter remains within the set normalization interval, thereby generating an initial entropy parameter that represents the disorder of user behavior.

[0058] By using the weighted summation processing method of the Shannon entropy aggregation mechanism described above, the intermediate information sequence of the previous step is transformed into a quantifiable and controllable initial entropy parameter, thereby achieving an overall measurement effect on the uncertainty of the evolution of user interests in a single window.

[0059] S2.4: Based on a preset entropy smoothing filtering method, perform noise reduction and correction processing on the initial entropy parameter to eliminate abnormal fluctuations caused by instantaneous operation jitter and generate a smooth behavioral activity entropy value with temporal stability.

[0060] The initial entropy parameter data output from step S2.3 is processed using a preset entropy smoothing filter method. The input includes an initial entropy value sequence indexed by time windows. A local average is calculated at the center of each time window based on a moving average operator; this local average serves as the basic filtering amount to remove random fluctuations. A Gaussian weighted function is used to apply a center weight to the initial entropy parameters within the window, with the weighting factor decaying exponentially based on the time offset from the window center, improving the ability to retain main trend information. An adaptive threshold detection logic is invoked to determine abnormal fluctuations in the difference between the moving average result and the original entropy parameters. If the difference exceeds a dynamic threshold, it is replaced with the local mean to suppress spikes. A second-order Butterworth low-pass filter is used to smooth the entropy value sequence after anomaly correction in the frequency domain; the filter cutoff frequency is dynamically set based on the user behavior sampling interval. The low-pass filter output and the moving Gaussian weighted average result are then weighted and fused. The combined weights are determined by the historical fluctuation stability coefficient and the density of behavioral events within the window, resulting in a smoothed behavioral activity entropy value with temporal stability.

[0061] By employing multi-stage smoothing and anomaly correction techniques, the initial entropy parameter from the previous step is transformed into a time-stable entropy value that eliminates instantaneous operational jitter while preserving the main trend change, thereby improving the accuracy and robustness of subsequent time-series difference analysis inputs.

[0062] S2.5: Encapsulate the smoothed behavior activity entropy value into a standardized data object to output the final single-window behavior activity entropy value that characterizes the uncertainty of user interest evolution within a single window, as the input benchmark for the subsequent time-series difference module.

[0063] like Figure 2 As shown, step S3 involves constructing a short time-series sequence from multiple consecutive single-window behavior activity entropy values ​​and inputting it into the time-series difference module. Dynamic boundary constraint operations are then performed to generate a personalized entropy threshold that drifts with the rhythm of individual user behavior. Specifically, this includes: S3.1: Obtain multiple consecutive single-window behavior activity entropy values ​​as input data, and perform temporal arrangement and alignment processing on the multiple consecutive single-window behavior activity entropy values ​​based on the sliding time stack mechanism to obtain a short temporal entropy sequence that characterizes the time distribution characteristics of the uncertainty of user interest evolution.

[0064] The standardized entropy data objects of multiple single-window behavioral activity values ​​output from the preceding step S2.5 are used as the initial input benchmark. The entropy data objects are sequentially pushed onto the timeline using a sliding time stack mechanism to ensure that the data from each window are arranged chronologically. Based on the entropy time series after the data is pushed onto the stack, window index mapping is performed to establish a one-to-one association between each entropy value and its corresponding timestamp, eliminating time misalignment caused by differences in collection batches. A timestamp alignment method is used to cross-reference entropy sequences from different batches, and missing or delayed window data is filled in to the expected time position through interpolation, resulting in a continuous and unbroken entropy sequence. The aligned entropy sequence undergoes time step size consistency processing, employing a periodic resampling strategy to unify the sampling step size of all sequences to a preset time interval, ensuring a consistent time benchmark for subsequent difference calculations. The continuous entropy sequence is constructed into a short-time entropy sequence data structure using data compression encoding. This data structure carries both timestamp index and entropy value sequence information tags during storage. By using a sliding time stack mechanism and a multi-batch alignment process, the entropy data from the previous step is transformed into a short temporal entropy sequence with a stable time base and continuous characteristics, thereby achieving an accurate representation of the temporal distribution characteristics of the uncertainty in the evolution of user interests.

[0065] S3.2: Receive the short-time entropy sequence as the processing object, and use the combination method of first-order difference operator and second-order difference operator to perform slope extraction and curvature analysis processing on the short-time entropy sequence to obtain the current window entropy growth rate representing the rate of change of user behavior activity and the entropy change acceleration vector representing the trend of change acceleration.

[0066] The short temporal entropy sequence generated by the preceding sub-step S3.1 is received as the input reference data for this sub-step; based on the differential operator initialization module, the input short temporal entropy sequence is mapped into a numerical matrix to ensure that the alignment accuracy between time indices meets the requirements of subsequent synchronous calculation.

[0067] The first-order difference operator is invoked to perform linear difference operations on the entropy values ​​of adjacent windows to extract the rate of change of entropy value at each time interval, forming the original vector of the current window's entropy growth rate. The following formula is used for the first-order difference calculation:

[0068] in, The entropy value corresponding to time index t. The entropy value corresponding to time index t-1, This represents the entropy growth rate corresponding to time index t.

[0069] Based on the first-order difference result, the second-order difference operator is called to perform linear difference operations on the first-order difference sequence to characterize the acceleration trend of the entropy change rate over time, generating the original vector of entropy-change acceleration; the second-order difference calculation formula is:

[0070] in, For entropy-change acceleration, This represents the entropy growth rate corresponding to time index t-1.

[0071] Normalization mapping is performed on the first-order and second-order difference results to ensure that the numerical range of entropy growth rate and entropy change acceleration is compatible with the input specifications of the subsequent dynamic threshold generation model; smoothing filtering is performed to remove instantaneous spikes or drops caused by abnormal operations of individual windows, thereby obtaining a stable current window entropy growth rate vector and entropy change acceleration vector.

[0072] By using differential combination calculation and data filtering, the short-time entropy sequence from the previous step is transformed into a two-parameter vector representing the rate of change and acceleration trend of user behavior activity, thus realizing the core time-series feature input required for dynamic boundary constraint calculation.

[0073] S3.3: Call the pre-stored historical behavior log database, and perform extreme value envelope fitting on the historical single-window behavior activity entropy value set based on the quantile regression method to obtain the lower bound entropy value curve and the upper bound entropy value curve that represent the stability boundary of users' long-term behavior, which are collectively referred to as the historical entropy fluctuation envelope. At the same time, perform weighted average calculation on the entropy benchmark dataset of the same user group based on the collaborative filtering clustering method to obtain the group entropy benchmark reference value that represents the normal level of group behavior.

[0074] Use the set of historical single-window behavior activity entropy values ​​recorded in chronological order in the pre-stored historical behavior log database as the input benchmark.

[0075] Based on this set of entropy values, the quantile regression method is called, and the target quantile set is set to 0.05 and 0.95, respectively, to fit the extreme trend curves of the low tail and the high tail, so as to ensure that the stability boundary of long-term user behavior is captured.

[0076] During the fitting process, the regression objective of each quantile is defined as the mapping relationship between the time window index and the corresponding entropy value, and a smoothing kernel function is used to locally weight the time series to reduce the impact of abnormal mutation points.

[0077] The low-tailed curve in the quantile regression fitting output is used as the lower bound entropy curve, and the high-tailed curve is used as the upper bound entropy curve. They are uniformly encapsulated as the historical entropy fluctuation envelope in the data structure.

[0078] We call the entropy benchmark dataset of similar user groups, and use the collaborative filtering clustering method to first group users according to the similarity of their behavior patterns. Then, we perform a weighted average calculation on the entropy value sequences of each group, where the weight coefficient is jointly defined by the number of samples in the group and the pattern similarity.

[0079] The output of the weighted average calculation is used to generate a baseline reference value for population entropy, ensuring that this reference value can characterize the normal level of population behavior and serve as an important input for subsequent dynamic boundary constraints.

[0080] Through the above processing method, the results of the previous step are transformed into a historical entropy fluctuation envelope containing upper and lower bound curves and a group entropy benchmark reference value, thereby achieving a dual characterization of the long-term stability of individual users and the normality of group behavior.

[0081] S3.4: Construct a dynamic boundary constraint operation model, taking the current window entropy growth rate, the entropy change acceleration vector, the historical entropy fluctuation envelope, and the group entropy benchmark reference value as multi-dimensional input variables, and performing multi-source feature fusion and nonlinear mapping transformation processing based on an adaptive weight allocation strategy to obtain an initial personalized entropy threshold candidate set that dynamically drifts over time and contains individual difference features.

[0082] Receive the current window entropy growth rate, entropy change acceleration vector, historical entropy fluctuation envelope, and group entropy benchmark reference value generated from the preceding sub-step S3.3 as multidimensional input variables.

[0083] The entropy growth rate and entropy change acceleration vector are input into the adaptive weight allocation strategy calculation unit. The behavioral fluctuation amplitude is calculated based on the difference between the upper and lower bound curves of the historical envelope, and the weight ratio of each input variable is adjusted accordingly.

[0084] The multi-source feature fusion module is invoked to concatenate the weighted entropy growth rate, entropy change acceleration vector, historical entropy fluctuation envelope, and population entropy benchmark reference value into a feature matrix in a multi-dimensional space, and principal component mapping is performed to reduce redundant correlations.

[0085] Based on the nonlinear mapping transformation mechanism, the fused feature matrix is ​​input into the radial basis function network, and Gaussian kernels are used to perform local sensitivity fitting on each component of the feature matrix to output the time drift response curve.

[0086] The dynamic boundary constraint formula is constructed using the intersection points of the curve peak and the envelope:

[0087] Where T is a single threshold in the initial personalized entropy threshold candidate set, S is the current window entropy growth rate, A is the entropy change acceleration measure, D is the historical entropy fluctuation envelope baseline value, and G is the group entropy baseline reference value. , , , Weighting coefficients for adaptive allocation.

[0088] The above formula is calculated point by point over a short time series to form an initial personalized entropy threshold candidate set that dynamically drifts over time and includes individual differences.

[0089] By using multi-source feature fusion and nonlinear mapping transformation, the results of the previous step are transformed into threshold candidate data that can be further smoothed and interpolated by S3.5, thus realizing the preliminary determination of personalized entropy threshold under dynamic boundary constraints.

[0090] S3.5: Obtain the initial personalized entropy threshold candidate set as the optimization object, and perform noise suppression and continuity correction processing on the initial personalized entropy threshold candidate set based on the smooth spline interpolation method to obtain the final personalized entropy threshold that drifts with the rhythm of individual user behavior and is used to determine behavioral pattern disturbance events.

[0091] like Figure 3 As shown, step S4 involves determining whether the single-window behavior activity entropy value of the current window exceeds the corresponding personalized entropy threshold. If the determination result is yes, the behavior pattern disturbance event identification process is triggered, and influence domain analysis is performed to obtain the set of keywords to be updated. Specifically, this includes: S4.1: Obtain the single-window behavior activity entropy value and the corresponding personalized entropy threshold of the current window, and use numerical comparison logic to perform size judgment processing on the two to obtain the behavior pattern disturbance event trigger signal that represents whether the behavior pattern has undergone drastic evolution.

[0092] S4.2: Receive the behavior pattern disturbance event trigger signal as an activation instruction, call the pre-constructed keyword co-occurrence graph data source, perform semantic coupling degree calculation processing based on the statistical results of high-frequency new words and long-tail words within the sliding time window, and obtain the candidate keyword adjacency matrix representing the strong correlation between words.

[0093] The behavior pattern perturbation event trigger signal generated by the preceding sub-step S4.1 is obtained as a data input condition. The trigger signal is sent as an activation instruction into the keyword co-occurrence graph calculation process, and the pre-constructed keyword co-occurrence graph data source that has been trained and updated over a long period of time is called.

[0094] Based on the sliding time window mechanism, a list of high-frequency new words and a list of long-tail words are extracted from the current window and its adjacent windows. Frequency statistics and occurrence position sampling are performed respectively to form a word occurrence frequency vector and a window position distribution vector.

[0095] Semantic embedding operations are performed on the list of high-frequency new words and the list of long-tail words. A predefined word vector model is called to obtain the corresponding high-dimensional semantic representation matrix, and the matrix is ​​normalized to eliminate the influence of vector length differences on subsequent calculations.

[0096] All word pairs and their semantic coupling degrees are arranged in matrix form. Threshold filtering logic is applied to remove word pairs with coupling degrees below the preset lower limit, while retaining word correspondences that conform to the association features.

[0097] By combining the word frequency vector, a weighted summation operation is performed on the retained word pair relationships. The semantic coupling degree is used as the first weight component and the frequency ratio is used as the second weight component. A comprehensive association weight is generated through linear weighting. The comprehensive association weight is filled into the corresponding cell positions of the candidate keyword adjacency matrix to form a candidate keyword adjacency matrix that represents the strong association relationship between words.

[0098] Through the above processing method, the behavioral pattern disturbance event trigger signal of the previous step is transformed into a candidate keyword adjacency matrix that can be used for subsequent path matching and quantitative evaluation of impact significance, thereby achieving high-precision lexical association data that integrates semantics and frequency.

[0099] S4.3: Obtain the recent click-jump path sequence data of the user, and use the graph traversal method to perform cross-class jump path matching processing on the click-jump path sequence and the adjacency matrix of the candidate keywords to extract the set of behavioral coupling paths that reflect the instantaneous transfer trajectory of user interests.

[0100] The recent click-through path sequence data from user interaction logs is obtained as the processing object. The path sequence is processed by temporal arrangement and index mapping based on timestamp order to ensure that the path nodes are completely aligned with the candidate keyword adjacency matrix output by the previous step S4.2 in both time and semantic dimensions.

[0101] The graph matching analysis engine is invoked to map the path sequence nodes one by one with the keyword nodes in the adjacency matrix, resulting in an initial matching graph structure containing node identifiers, edge weights, and path category labels.

[0102] In the initial matching graph structure, a depth-first traversal method is applied to perform a cross-category node combination search for each click jump path, identify the complete edge link from one type of keyword node to another type of keyword node, and mark its cross-category jump characteristics.

[0103] A path pattern filtering mechanism is applied to the detected cross-class transition paths to remove noisy paths that do not meet the minimum path length or minimum edge weight threshold, and retain a significant subset of paths that simultaneously possess high-frequency access characteristics and keyword category span characteristics.

[0104] The significant path subset is encapsulated into a behavior-coupled path set data object, and path access frequency, node category sequence and edge weight distribution are added as extended attributes. Through cross-class transition matching processing, the candidate keyword adjacency matrix of the previous step is transformed into a behavior-coupled path set with the instantaneous transfer trajectory characteristics of user interests, so as to achieve accurate characterization of the influence domain of disturbance events.

[0105] S4.4: Based on the node access frequency and edge weight distribution in the set of behavioral coupling paths, the centrality ranking method is used to perform quantitative evaluation of the influence significance of candidate keywords, and obtain the keyword influence score vector that represents the degree of influence of each candidate keyword on the current behavioral disturbance.

[0106] The behavior coupling path set generated by the previous sub-steps is obtained as the input data object. The node access frequency analysis module is called to perform access frequency statistics calculation on each node in the behavior coupling path set, obtaining an access intensity index representing the access intensity of each candidate keyword in the current perturbation scenario. Based on the access intensity index, the edge weight distribution analysis module is called to perform weight extraction processing on each edge in the behavior coupling path set, using the weight values ​​to reflect the semantic and behavioral coupling tightness between keywords, and forming a cross-node association strength matrix. The access intensity index and association strength matrix are used as dual input variables, and the degree centrality and weighted adjacency calculation sub-modules in the centrality ranking method are loaded to perform deduplication and ranking processing on the comprehensive influence significance score, generating a keyword influence score vector arranged from high to low influence degree. By combining centrality ranking with access and weight dual-factor processing, the coupling path data of the previous step is transformed into a quantitative score vector that can be used for perturbation-sensitive keyword screening, realizing accurate positioning of the influence domain in high perturbation scenarios.

[0107] S4.5: Read the keyword influence score vector and apply a dynamic truncation threshold filtering strategy. Perform set encapsulation processing on candidate keywords with scores higher than the threshold to obtain the final set of keywords to be updated for guiding the fine-tuning of local model parameters.

[0108] The keyword influence score vector data object generated in the previous steps is read, and a dynamic truncation threshold calculation model adapted to the current behavioral disturbance intensity is constructed based on the numerical distribution characteristics of the vector elements. A combination of moving average and standard deviation is used to extract statistical features from the influence score vector, forming mean and fluctuation range parameters to characterize the overall disturbance level, which serve as the input benchmark for dynamic threshold calculation. Based on the extracted mean and fluctuation range parameters, an initial truncation threshold candidate set is generated using a linear offset rule, where the offset factor is driven by the normalization coefficient of the current window behavior activity entropy value to ensure that the threshold adaptively adjusts with the disturbance intensity. Smoothing filtering and boundary correction processing are performed on the initial truncation threshold candidate set to eliminate extreme abnormal thresholds and limit their drift within a preset operable range, resulting in the final executable dynamic truncation threshold. The dynamic truncation threshold is read and applied to the keyword influence score vector, and a filtering operation based on numerical comparison logic is performed to encapsulate candidate keywords with scores higher than the threshold into a set data object according to their original sequence position index. The data objects in the set are appended with structured metadata, including timestamps, perturbation event identifiers, and the high-entropy window number to which they belong, to form the final set of keywords to be updated, which is used to guide the fine-tuning of local model parameters. Through a dynamic truncation threshold filtering strategy, the impact score results from the previous step are transformed into a structured set of keywords to be updated, achieving precise limitation of the local model input domain and resource-sensitive update control.

[0109] Step S5: Construct a lightweight trend modeling submodule consisting of gated linear units superimposed with a residual attention mechanism, and use the set of keywords to be updated as input to perform local gradient correction. Specifically, this includes: S5.1: Obtain a predefined lightweight neural network architecture template, perform layer stacking processing based on the gated linear unit method and the residual attention mechanism method to obtain a lightweight trend modeling sub-module topology structure containing an input layer, a hidden layer and an output layer. The hidden layer integrates a gating matrix for controlling information flow and a residual attention weight matrix for capturing long-range dependencies.

[0110] The basic structural information, consisting of a predefined lightweight neural network architecture template, is obtained, specifying the number of nodes, data format, and connection methods for the input, hidden, and output layers. Based on the architecture template, a gating matrix for controlling information flow is constructed using the gated linear unit method. Control signal transmission paths and state preservation mechanisms are defined in the hidden layer, and the activation conditions and initial bias values ​​for each gating channel are set to ensure efficient data selection and state updates under resource constraints. A residual attention mechanism is introduced into the hidden layer, establishing a weight matrix for residual connections. The attention level of each node to long-range dependencies is allocated through weight position mapping rules, and a multi-head attention module is embedded in the structure to enhance the ability to capture associations across different feature dimensions. Using hierarchical stacking, the gating matrix and the residual attention weight matrix are combined and embedded into the hidden layer units in a predetermined hierarchical order. This ensures that information flow is gated during forward propagation while preserving contextual features propagated back through the residual path, avoiding gradient vanishing and enhancing local sensitivity. Structural consistency checks and parameter constraint processing are performed on the connections between the input layer and hidden layers, and between the hidden layers and the output layer, to ensure that the overall topology meets the design goals of low latency and fast response of the lightweight trend modeling submodule. Through this processing, the access conditions of the keyword set to be updated in the previous step are transformed into a data stream that can be stably transmitted and dynamically adjusted within the lightweight trend modeling submodule. This enables the model to maintain structural stability while quickly sensing behavioral disturbances and providing a high-precision input environment for local gradient correction.

[0111] S5.2: Read the historical score backbone parameter data corresponding to the keyword set to be updated, and fill the historical score backbone parameter data into the input layer node of the lightweight trend modeling submodule topology based on the parameter mapping rules to obtain the initial state tensor carrying the historical interest evolution characteristics. This initial state tensor serves as the benchmark reference system for subsequent local correction operations.

[0112] Read the historical score backbone parameter data corresponding to the keyword set to be updated, call the structured data interface to obtain the original numerical matrix of time decay coefficient, interest migration weight and context stability factor for each keyword, and perform index matching according to the keyword unique identifier to ensure that the parameter data corresponds one-to-one with the target set.

[0113] Based on the parameter mapping rules, the time decay coefficient, interest migration weight, and context stability factor are mapped to the specified nodes of the topology input layer of the lightweight trend modeling submodule. Each type of parameter is normalized and transformed into dimensionless feature values ​​to adapt to the uniform scale requirements of the network input layer.

[0114] The normalized parameter data is subjected to feature concatenation operation, which merges the three types of parameters of the same keyword into a single vector to form a high-dimensional input feature array containing static historical interest evolution information. The feature vectors of all keywords are batch-encapsulated through matrix rearrangement process to meet the batch input format of the model.

[0115] The initial state tensor is generated in the computing environment using the tensor construction method. The batch-encapsulated feature array is mapped into a three-dimensional tensor structure that conforms to the dimensional specifications of the lightweight modeling sub-module. The first dimension represents the keyword sample index, the second dimension represents the parameter category index, and the third dimension represents the normalized feature value sequence.

[0116] The tensor verification mechanism checks the integrity and data type consistency of the initial state tensor, ensuring that the historical scoring backbone parameters corresponding to each keyword have been correctly filled into the input layer nodes, providing a reliable benchmark reference for subsequent local gradient correction operations.

[0117] Through the above parameter mapping and tensor construction processing, the set of keywords to be updated in the previous step and its historical score backbone parameters are transformed into initial state tensor data that carries the historical interest evolution characteristics, so as to realize the effective carrying and structural matching of the input by the lightweight trend modeling submodule before local correction.

[0118] S5.3: Receive all network weight parameters in the topology of the lightweight trend modeling submodule, and perform locking processing on all network weight parameters except for the time decay coefficient, interest migration weight and context stability factor based on the backbone network freezing strategy, to obtain a local gradient correction enabled model that only contains fine-tunable parameter channels, ensuring that the historical score backbone parameters remain unchanged during the update process.

[0119] It receives all network weight parameters from the topology of the lightweight trend modeling submodule, which includes the input layer, hidden layer, and output layer. It then classifies and labels the weight matrix according to its functional channel to distinguish the corresponding channel and non-target channel of the time decay coefficient, interest migration weight, and context stability factor.

[0120] The backbone network freezing strategy module is invoked to perform numerical locking on the weight matrix of non-target channels, disconnecting their gradient calculation paths and placing them into the freeze mask matrix to ensure that such weights remain fixed in subsequent optimization stages.

[0121] Based on the weighted classification labeling results, a channel index table with finely adjustable parameters is constructed. Only the matrix position indexes of the time decay coefficient, interest migration weight, and context stability factor are retained in this table as the sole entry point for the local update path.

[0122] The submodule weight parameters are masked using a fine-tunable channel index table. The weight values ​​at non-indexed positions are combined with the frozen mask matrix and written into the weight locking buffer. The weight values ​​at indexed positions are kept in the active state of gradient updates, so as to form a local gradient correction enabled model that only contains fine-tunable parameter channels.

[0123] A consistency check method is used to compare the numerical consistency between the frozen mask matrix and the weight locking buffer to ensure that the historical scoring backbone parameters have not been updated incorrectly. A fully functional gradient correction enabled model containing only locally correctable parameters is generated through structured output to ensure the numerical fidelity of the historical scoring backbone parameters during the model update process.

[0124] By combining the backbone network freezing strategy with the local gradient correction enabled model, the initial state tensor of the previous step is transformed into a network structure with fine-tuning capabilities for a specific set of parameters, thereby achieving a resource-sensitive high-precision local response effect.

[0125] S5.4: Call the gated linear unit component inside the local gradient correction enabling model, and perform forward propagation derivation based on the initial state tensor to obtain an intermediate activation vector that represents the temporal features of the current keyword. This intermediate activation vector integrates the static features of the historical scoring backbone parameters and the dynamic context information of the current time window.

[0126] The gated linear unit (LMU) component within the locally gradient-corrected enabling model, which has had its non-tunable network weights locked and adjustable parameter channels preserved in step S5.3, is obtained. The initial state tensor generated in step S5.2 is loaded as the input signal matrix of this component, ensuring that the parameter mapping relationship between the input layer and hidden layer channels is arranged in a consistent topological structure. A single-step matrix multiplication operation is performed on the loaded initial state tensor, performing a dot product between the input signal and the weight coefficient matrix of the gate matrix in the corresponding row and column spaces. Subsequently, a bias vector addition operation is applied to the product matrix to form a linear combination result without nonlinear transformation. A preset activation function operator is invoked to perform element-wise nonlinear mapping processing on the linear combination result, mapping the values ​​to a limited range to enhance the nonlinear expressive power of the network. The activation function can be a hyperbolic tangent or rectified linear unit. Based on the obtained nonlinear transformation result, an element-wise multiplication operation of the gating mechanism is performed, calculating the Hadamard product between the activation value and the corresponding gate switch value matrix to filter out low-correlation or low-sensitivity feature components, thereby obtaining an intermediate state matrix with time-series feature enhancement. The intermediate state matrix is ​​normalized to eliminate scale differences between different feature channels, outputting an intermediate activation vector representing the temporal features of the current keyword. This vector integrates the static features of the historical scoring backbone parameters with the dynamic context information of the current time window. Through the matrix multiplication, nonlinear mapping, gated multiplication, and normalization processes described above, the initial state tensor from the previous step is transformed into an intermediate activation vector containing temporal sensitivity and context fusion characteristics, thus preparing the input for local gradient correction.

[0127] S5.5: The intermediate activation vector is weighted and aggregated using the residual attention weight matrix. The aggregation result is fused with the original input features based on the residual connection mechanism to obtain the final model ready state with enhanced local sensitivity. This final model ready state indicates that the lightweight trend modeling submodule is ready to receive the entropy constraint signal and perform specific gradient correction operations.

[0128] The residual attention weight matrix is ​​called to perform matrix multiplication on the intermediate activation vector generated by the gated linear unit to form a weighted context-sensitive response vector that retains the original temporal features and introduces cross-window dependencies.

[0129] The response vector is normalized, and the weighted values ​​of each dimension are mapped to a preset numerical range to eliminate the bias of the weighting effect at different scales and ensure the stability of subsequent fusion.

[0130] Based on the residual connection mechanism, the normalized response vector is added element-wise to the original input feature tensor to construct a fusion feature matrix that includes static behavioral features and dynamic context compensation.

[0131] The consistency score of the fused feature matrix is ​​scored using a feature consistency test method. If the score is lower than a preset threshold, a smoothing correction process is triggered. The weights of feature components with large differences in the matrix are redistributed to enhance local sensitivity.

[0132] The fusion feature matrix, after consistency verification and correction, is encapsulated into a final model-ready state data object. This data object has the dual ability to retain historical scoring backbone parameters and adapt to current window context perturbations, and can be directly used as input for entropy constraint signals to drive local gradient correction.

[0133] By using residual attention weight matrix weighted aggregation and residual fusion processing, the intermediate activation vector from the previous step is transformed into a model-ready state with enhanced local sensitivity, thereby achieving high response performance of the lightweight trend modeling submodule to dynamic disturbances.

[0134] Step S6: Based on the linear constraint of the magnitude of the single-window behavior activity entropy value, the adjustment range is modified. The lightweight trend modeling submodule performs local gradient correction operations only on the set of keywords to be updated, generating keyword scoring results. Based on these keyword scoring results, the prediction of potential keywords is completed. Specifically, this includes: S6.1: Obtain the single-window behavior activity entropy value and the preset entropy value normalization interval, and use the linear mapping function to perform scaling transformation on the single-window behavior activity entropy value to obtain the entropy-driven correction coefficient that characterizes the intensity of the current behavior disturbance. This entropy-driven correction coefficient will be used as the global control variable for the subsequent parameter adjustment range.

[0135] S6.2: Read the historical score backbone parameters of each candidate keyword in the keyword set to be updated, and analyze the time dimension features in the historical score backbone parameters based on the timestamp sequence to extract the initial value of the time decay coefficient to be corrected, which includes the original time decay rate, and provide a benchmark reference for subsequent gradient correction.

[0136] The process reads historical rating backbone parameter data from the keyword set to be updated as the execution object. Based on the timestamp sequence parsing module, it performs structured separation processing on the time dimension features of the historical rating backbone parameters to extract time decay-related components that are separate from rating weights and context factors. A time series synchronizer is invoked to arrange the time dimension features in ascending order and fill in missing sequence nodes, ensuring the continuity and integrity of the time decay data. A sliding regression window mechanism is used to perform local fitting processing on the time decay-related components, generating a decay curve coefficient matrix that characterizes the natural decay trend of user interest over time. Based on an exponential fitting function, parameters are extracted from the decay curve coefficient matrix, and the fitted coefficients are converted into initial decay rate values. A linear normalization module is invoked to normalize the calculated initial decay coefficient values ​​across the entire keyword set to be updated, forming a unified-scale decay coefficient reference set for subsequent gradient correction. Through this normalized decay coefficient reference set, the results of the previous step are transformed into initial time decay coefficient data that can directly drive the lightweight trend modeling submodule, achieving the expected technical effect of providing a consistent benchmark for gradient correction.

[0137] S6.3: Call the lightweight trend modeling submodule, which consists of gated linear units superimposed with residual attention mechanism, and use the initial value of the time decay coefficient to be corrected, the semantic vector of the keyword set to be updated, and the context features as input data to perform forward propagation calculation, so as to output the interest migration gradient vector representing the instantaneous shift of user interest and the context stability gradient vector representing the degree of environmental noise interference.

[0138] The initial value of the time decay coefficient to be corrected is used as the input of the time series component, and together with the semantic vector and context features of the keyword set to be updated, it is loaded in parallel to the input layer node of the lightweight trend modeling submodule.

[0139] The above input is controlled dimensionally by gated linear unit matrix. By selectively suppressing time decay coefficient components that do not have high entropy sensitivity through gating function, the semantic and contextual components that are strongly correlated with the current user interest disturbance are strengthened.

[0140] By using the residual attention weight matrix to capture long-range dependent features on the intermediate state vector after gating, the temporal decay trend and cross-window interest transfer pattern are weighted and fused in the attention space to form a high-order feature set containing the directionality of interest transfer.

[0141] The forward propagation computation method is used to perform node activation operations on the fused high-order feature set. The gradient signal is transmitted between network layers through a nonlinear activation function, forming two independent output channels. The interest transfer gradient vector is used to characterize the instantaneous shift of interest within a short window, while the context stability gradient vector is used to quantify the degree of environmental noise and context fluctuation.

[0142] The interest transfer gradient vector and the context stability gradient vector are encapsulated with dimension and type labels for subsequent entropy-driven correction coefficient weighted fusion processing.

[0143] Through the above processing method, the initial value of the time decay coefficient, semantic vector and context features in the previous step are transformed into interest migration gradient vector and context stability gradient vector that can be directly used for parameter fine-tuning, so as to realize the dynamic scoring model's ability to perceive user interest shifts and environmental disturbances in real time.

[0144] S6.4: Based on the entropy-driven correction coefficient, the interest migration gradient vector and the context stability gradient vector are weighted and fused. The dynamic step size control method is used to truncate and smooth the fused gradient data to obtain the time decay correction increment for the time decay coefficient, the interest migration correction increment for the interest migration weight, and the context stability correction increment for the context stability factor.

[0145] Obtain the interest migration gradient vector and context stability gradient vector output from the preceding sub-steps, and import the entropy-driven correction coefficient as a global weight control variable into the fusion calculation logic.

[0146] The weighted fusion model is initialized based on the weight coefficients. The interest transfer gradient vector and the context stability gradient vector are multiplied and accumulated in the same fusion space to obtain the intermediate matrix of the fusion gradient.

[0147] The dynamic step size control method is invoked to determine the step size adjustment function curve based on the real-time value of the entropy-driven correction coefficient, and to perform step size adjustment operations on each element of the fused gradient intermediate matrix, so that the gradient change amplitude and the entropy disturbance intensity form a linear mapping relationship.

[0148] A truncation mechanism is adopted to perform threshold truncation operation on elements in the fused gradient data after step size adjustment that exceed the preset safety threshold, so as to ensure that the parameter correction increment does not exceed the limit.

[0149] The smoothing module is invoked, and the bidirectional moving average filtering method is used to perform smoothing correction on the truncated fused gradient data, eliminating the impact of abrupt changes on the stability of local parameter updates.

[0150] Based on the smoothed fused gradient data, the corresponding components of the time decay coefficient, interest migration weight, and context stability factor are extracted to generate time decay correction increment, interest migration correction increment, and context stability correction increment, respectively.

[0151] Through the above-mentioned fusion, step size control, truncation and smoothing methods, the gradient vector of the previous step is transformed into three executable parameter correction increments for different keyword temporal features, thereby achieving precise dynamic adjustment of the local model.

[0152] S6.5: The time decay correction increment, interest transfer correction increment, and context stability correction increment are superimposed onto the corresponding initial values ​​of the time decay coefficient, interest transfer weight, and context stability factor, respectively. The parameter recalibration aggregation operation is then performed to obtain the dynamically adjusted keyword scoring results with time-adaptive characteristics and complete this local model update.

[0153] Step S7: Record update log data and inject the update log data as feedback into the data source. Specifically, this includes: S7.1: Obtain the original update event data containing the trigger timestamp, the set of keywords to be updated, and the magnitude of parameter adjustment. Use the time-series alignment method to standardize and encapsulate the original update event data to obtain the initial record of the structured update log with a unified spatiotemporal benchmark.

[0154] S7.2: Based on the trigger timestamp in the initial record of the structured update log, call the preset time window sliding retrieval interface to perform a range query operation on the historical clickstream database to extract the original sequence of user interaction behavior within the preset time span corresponding to the trigger timestamp.

[0155] Based on the trigger timestamp in the initial record of the structured update log, the initial condition for performing a range query operation on the historical clickstream database by calling the preset time window sliding retrieval interface is that the trigger timestamp has been standardized and encapsulated through step S7.1 and has a unified spatiotemporal reference index. The retrieval execution object is the user interaction event sequence in the historical clickstream database, which includes the click target identifier, event occurrence time, session identifier, and jump path data. The target retrieval window parameters centered on the trigger timestamp are initialized using the time window sliding retrieval interface, including the window start time and window end time. The window span is jointly calculated based on the preset time span configuration value and the time zone offset of the trigger timestamp. The index positioning method is called to lock the start and end boundary positions of the window in the historical clickstream database, and the user interaction event records that meet the time constraints are extracted between the boundaries through a bidirectional pointer traversal mechanism. The order of the extracted original interaction events is rearranged using the event timestamp to ensure the continuity and correct order of the time sequence. The event type filtering module is executed to remove non-click categories and interaction behaviors unrelated to keyword scoring adjustment, and retain valid events with keyword matching items or jump path associations. The data integration module encapsulates the fields of valid events into a raw sequence object of user interaction behavior in a unified format. This object contains event time, keyword identifier, jump path and session context information.

[0156] Through the above-mentioned sliding retrieval and time boundary control processing method, the initial record of the structured update log in the previous step is transformed into the original sequence of user interaction behavior within a preset time span corresponding to the trigger timestamp, so as to provide a continuous, complete and time-aligned input benchmark for S7.3 multidimensional attribution analysis.

[0157] For example, in a search engine optimization (SEO) scenario, when the trigger timestamp in the initial record of the structured update log is 14:35:27 on May 12, 2024, the preset time span configuration value is set to 7200 seconds, the time zone offset of the trigger timestamp is +08:00, the starting time of the sliding search interface calculation window is 13:35:27 on May 12, 2024, and the ending time of the window is 15:35:27 on May 12, 2024, the historical clickstream database is scanned using this time interval as the boundary. All interaction events of user IDs U102, U356, and U784 within this time range are retrieved through index positioning, resulting in 148 records. After the event type filtering module removes non-click events such as browsing, scrolling, and form submissions, 38 valid click event records containing any keyword from the set of keywords to be updated are retained. These 38 records were sorted by event timestamps and encapsulated into a raw sequence object of user interaction behavior. Each record included keyword identifiers (e.g., K114, K223), jump paths (e.g., / ads / product / 322), session identifiers, and the context state of the preceding and following sessions. This raw sequence was used in subsequent step S7.3 for conversion path mapping and calculation of click-through rate improvement. Validation results showed that the temporal continuity and keyword matching accuracy of this sequence were significantly improved, providing high-quality data input for attribution analysis.

[0158] S7.3: Perform multidimensional attribution analysis on the original sequence of user interaction behavior, and use the conversion path mapping method to identify click conversion events directly driven by the set of keywords to be updated, so as to calculate the click conversion improvement rate, which represents the actual business value of this rating adjustment.

[0159] The system receives the original sequence of user interaction behaviors within a preset time span corresponding to the trigger timestamp as the analysis object, calls the multidimensional attribution analysis processing module, performs keyword binding mapping on each interaction event in the original sequence of user interaction behaviors, performs one-to-one or one-to-many matching verification between the text keywords in the event and the set of keywords to be updated, and generates an event label sequence containing keyword matching tags.

[0160] A transformation path mapping method is performed on the event annotation sequence to construct a directed graph of click paths based on the user session link sequence. Each event node is connected in chronological order, and the jump relationship and edge weight information between nodes are retained to form a set of directed paths for path analysis.

[0161] In the set of directed paths, path links whose starting point contains the set of keywords to be updated and whose ending point is a conversion event are identified. Path filtering rules are used to remove links that do not conform to the direct driving relationship from keywords to conversion, thus obtaining a set of effective conversion paths directly driven by the set of keywords to be updated.

[0162] Perform event statistics processing on the set of effective conversion paths, calculate the number of conversion events in the path as the number of conversion events directly driven by the set of keywords to be updated, and call the historical benchmark data interface to obtain the total number of conversion events within the corresponding time window.

[0163] Using the number of conversion events directly driven by the set of keywords to be updated and the total number of conversion events as input, the click-to-conversion rate improvement formula is used:

[0164] Where C represents the number of conversion events directly driven by the set of keywords to be updated in the current window. The baseline number of conversion events directly driven by the set of keywords to be updated in the previous period is used to perform numerical calculations to obtain the click-to-conversion improvement rate, which represents the actual business value of this score adjustment.

[0165] By using multidimensional attribution analysis and conversion path mapping methods, the original sequence of user interaction behavior in the previous step is transformed into quantifiable click-to-conversion rate data, enabling a direct correlation assessment between the effect of local model updates and business benefits.

[0166] S7.4: The click-to-conversion rate value is backfilled into the initial record of the structured update log. The trigger timestamp, the set of keywords to be updated, the parameter adjustment level, and the click-to-conversion rate value are integrated into a complete closed-loop feedback data package using a data fusion protocol to form an update log data entity containing a causal chain.

[0167] The click-through rate improvement value is imported into the initial record storage module of the structured update log as a new input variable. The index matching logic is called to associate it with the trigger timestamp, the set of keywords to be updated, and the corresponding fields of parameter adjustment level, so as to build a log augmentation dataset with multi-dimensional data binding relationship.

[0168] In the log augmentation dataset, joint validation of numerical and temporal data is performed based on field type matching rules to ensure that the click conversion rate and the trigger timestamp are causally traceable within a preset time window, thereby providing a reliable data source for the generation of subsequent feedback packages.

[0169] By utilizing the multi-field concatenation function of the data fusion protocol framework, the trigger timestamp, the set of keywords to be updated, the parameter adjustment level, and the click-through rate improvement value are combined into a continuous binary data segment according to the preset field sequence rules. Field identifier bits are injected into the protocol header to indicate the logical position of each parameter in the feedback packet.

[0170] The protocol stack's encapsulation interface is invoked to fully encapsulate the binary data segment and protocol header information, generating a closed-loop feedback data packet containing metadata of the causal chain. This data packet contains full-process mapping information from behavior triggering to business benefits.

[0171] Perform a consistent hash check on the closed-loop feedback data packets to verify the integrity of the encapsulation and the immutability of the data. Output the data packets that pass the check as update log data entities.

[0172] By using data fusion protocols and encapsulation processing methods, the click-to-conversion rate improvement results from the previous step and the original update log records are transformed into update log data entities containing causal chains, thereby realizing structured feedback output that directly links local model update behavior with business results.

[0173] For example, in the search promotion system, the trigger timestamp field is configured as 2024-05-12 14:35:26, the set of keywords to be updated is set to 18, the parameter adjustment level is 0.025, and the click-through rate improvement rate is calculated to be 1.42 using the conversion path mapping method. These four fields are concatenated into a continuous binary data segment using a data fusion protocol. The trigger timestamp occupies 8 bytes, the set of keywords to be updated occupies 4 bytes, the parameter adjustment level occupies 4 bytes, and the click-through rate improvement rate occupies 4 bytes. The protocol header contains four field identifiers and version number information, occupying 6 bytes, for a total encapsulation length of 30 bytes. During the field association verification process, the system confirms that the interval between the click-through rate improvement rate and the trigger timestamp does not exceed a preset 36-hour window, satisfying the causal tracing condition. Subsequently, the protocol encapsulation interface is called to generate a closed-loop feedback data packet, and a consistent hash verification is performed on the data packet. A successful hash value match is obtained, ensuring that the data packet is not tampered with during transmission. The final output update log data entity contains end-to-end mapping information from triggering behavior to business benefits. In the deployment environment, it was verified that this entity can be directly used as input to the threshold optimization module. In multiple batch tests, it showed a significant improvement in model response performance and system prediction accuracy.

[0174] S7.5: Perform serialization encoding processing based on the update log data entity, and use the message queue push mechanism to inject the encoded update log data entity into the feedback injection data source storage area to complete the data flow preparation from the local model response effect to the global threshold optimization control signal.

[0175] Step S8: Perform iterative optimization processing on the generation logic of the personalized entropy threshold based on the feedback injection data source. Specifically, this includes: S8.1: Obtain update log data containing trigger timestamps, set of keywords to be updated, parameter adjustment magnitude, and click-through rate improvement within a preset time after the update. Perform time-series alignment and outlier removal on the update log data to obtain a standardized effect feedback sequence as the input basis for subsequent analysis.

[0176] S8.2: Based on the click-to-conversion rate improvement index in the effect feedback sequence, perform a correlation mapping operation to construct an effectiveness evaluation vector that represents the causal relationship between the score update operation and the actual business revenue, so as to quantify the effectiveness of the model response under different entropy trigger intervals.

[0177] S8.3: Using the performance evaluation vector, perform gradient backpropagation optimization on the weight allocation coefficients among the historical entropy fluctuation envelope, the current window entropy growth rate, and the entropy benchmark of the same user group to obtain a dynamic boundary constraint parameter set that can maximize long-term click conversion revenue.

[0178] The performance evaluation vector generated by the preceding step S8.2 is used as the optimization input condition, and each component in the vector is regarded as a quantitative indicator of the effectiveness of the model response under different entropy triggering intervals.

[0179] Read the stored historical entropy fluctuation envelope curve data, the current window entropy growth rate value, and the entropy benchmark reference value of the same user group, and input the three as dependent variables of the dynamic boundary constraint parameter weight allocation coefficient to be optimized into the optimization model.

[0180] A gradient backpropagation optimization framework is constructed, the weight allocation coefficients are initialized to preset values, and the long-term click-to-conversion revenue maximization objective is embedded in the loss function. It can be represented as:

[0181] in This represents the long-term click-to-conversion revenue forecast; the negative sign is used to maximize the target.

[0182] When performing forward propagation, the predicted click-to-conversion revenue value R is generated by weighted fusion of the historical entropy fluctuation envelope and the group entropy benchmark, and the deviation between the predicted value and the actual revenue value in the performance evaluation vector is calculated.

[0183] The gradient calculation operator is invoked to calculate the partial derivatives of the weight allocation coefficients with respect to the loss function L based on the chain rule. An example of the gradient formula is shown below:

[0184] in The weighting coefficients for the historical entropy fluctuation envelope.

[0185] The gradient values ​​are input into the weight update operator, and the learning rate decay strategy is used to adjust the values ​​of each weight allocation coefficient. After the update, the weight allocation coefficients flow back to the input of the optimization model, forming an iterative optimization closed loop.

[0186] By using the gradient backpropagation optimization method, the performance evaluation vector from the previous step is transformed into a dynamic boundary constraint parameter set that can maximize long-term click-to-conversion revenue, thereby achieving adaptive tuning of the entropy threshold generation logic.

[0187] S8.4: Based on the dynamic boundary constraint parameter set, perform adaptive reconstruction processing on the drift rate and convergence interval of the personalized entropy threshold to obtain the next generation personalized entropy threshold generation logic rule with self-evolution capability.

[0188] Based on the generated set of dynamic boundary constraint parameters, the weight allocation coefficients for the historical entropy fluctuation envelope, the current window entropy growth rate, and the group entropy benchmark are read and combined with the statistical characteristics of the current short-time entropy sequence as input variables for adaptive reconstruction.

[0189] The drift rate calculation unit is invoked to perform a weighted summation of the gradient of the personalized entropy threshold within a continuous time window. The weight coefficients used are provided by the dynamic boundary constraint parameter set to quantify the dynamic offset intensity of the threshold over time.

[0190] By using the convergence interval analysis module, the minimum and maximum values ​​of the threshold in the past multiple iterations are used to form the interval boundary. Based on the stability factor in the dynamic boundary constraint parameter set, the interval width is adjusted to ensure that the threshold has the ability to expand under high volatility behavior mode and the ability to contract under low volatility behavior mode.

[0191] A nonlinear mapping function is used to integrate the drift rate and convergence interval as dual indices. The output of the evolution function is then input into the smooth spline reconstruction module to perform continuity correction and noise suppression processing, so that the reconstructed threshold curve remains stable in the time series and has self-adjustment capability.

[0192] By using an adaptive reconstruction method, the dynamic boundary constraint parameter set is transformed into a next-generation personalized entropy threshold generation logic rule with time drift adaptability and interval stability, thereby realizing the self-evolution function of the threshold calculation unit.

[0193] S8.5: Deploy the next-generation personalized entropy threshold generation logic rule to the threshold calculation unit of the time-series difference module, and complete the hot replacement operation of the old generation logic to formally establish the control mechanism of behavior uncertainty perception, local model response and effect feedback threshold closed-loop evolution.

[0194] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0195] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0196] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for mining and predicting potential keywords based on time-series search behavior analysis, characterized in that, Specifically, it includes: S1: Obtain user search behavior stream data, and slice the search behavior stream data based on the sliding time window mechanism to generate a four-dimensional indicator behavior activity vector; S2: Use the information entropy calculation formula to perform discrete quantization on the behavior activity vector to obtain the single-window behavior activity entropy value; S3: Construct a short time series from multiple consecutive single-window behavior activity entropy values ​​and input it into the time series difference module to perform dynamic boundary constraint operation and generate a personalized entropy threshold that drifts with the rhythm of individual user behavior. S4: Determine whether the single-window behavior activity entropy value of the current window exceeds the corresponding personalized entropy threshold. If the determination result is that it exceeds the threshold, trigger the behavior pattern disturbance event identification process and perform influence domain analysis to obtain the set of keywords to be updated. S5: Construct a lightweight trend modeling submodule consisting of gated linear units superimposed with a residual attention mechanism, and take the set of keywords to be updated as input to perform local gradient correction; S6: Based on the linear constraint of the magnitude of the single-window behavior activity entropy value, the lightweight trend modeling submodule performs local gradient correction operation only on the set of keywords to be updated to generate keyword scoring results, and completes the prediction of potential keywords based on the keyword scoring results.

2. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 1, characterized in that, The process following step S6 also includes: S7: Record update log data and inject the update log data as feedback into the data source; S8: Perform iterative optimization processing on the generation logic of the personalized entropy threshold based on the feedback injection data source.

3. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 1, characterized in that, The behavioral activity vector of the four-dimensional indicator includes four types of vectors: keyword frequency, behavioral type diversity, session interval distribution, and cross-device consistency.

4. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 1, characterized in that, The temporal difference module specifically performs temporal regularization on the single-window behavior activity entropy value using a sliding time stack, extracts the entropy growth rate and entropy change acceleration vectors through first- and second-order difference operators, and analyzes the rate and trend of change in single-window behavior activity; combines quantile regression to fit the historical entropy fluctuation envelope and collaborative filtering clustering to obtain the group entropy benchmark reference value, and finally optimizes and corrects it through smooth spline interpolation to output a personalized entropy threshold that adapts to the user's behavior rhythm and can determine the behavior pattern disturbance event.

5. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 2, characterized in that, The update log data includes the trigger timestamp of the sliding time window and the set of keywords to be updated.

6. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 1, characterized in that, Step S2 specifically includes: Obtain the behavior activity vector, and perform normalization mapping on the original values ​​of each dimension in the behavior activity vector to generate standardized behavior probability distribution basis data; Based on the standardized behavioral probability distribution baseline data, a negative logarithmic weighting operation is performed on each dimension component using a logarithmic transformation method to map the linear probability space to the information metric space and generate an intermediate information quantity sequence containing the independent information contribution of each dimension. The intermediate information sequence is weighted and summed using the Shannon entropy aggregation mechanism to generate an initial entropy parameter that represents the disorder of user behavior. The initial entropy parameter is denoised and corrected using an entropy smoothing filtering method to generate a smooth behavioral activity entropy value with temporal stability. The smoothed behavior activity entropy value is encapsulated into a standardized data object to output the final single-window behavior activity entropy value that characterizes the uncertainty of user interest evolution within a single window.

7. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 1, characterized in that, Step S4 specifically includes: The trigger signal for behavior pattern disturbance events is calculated based on the single-window behavior activity entropy value and the personalized entropy threshold. The adjacency matrix of candidate keywords is obtained by calculating and processing the trigger signals of behavioral pattern disturbance events. A set of behavior-coupled paths is obtained by path matching based on the adjacency matrix of candidate keywords; Based on the node access frequency and edge weight distribution in the set of behavioral coupling paths, the centrality ranking method is used to perform quantitative evaluation of the influence significance of candidate keywords, and obtain the keyword influence score vector that represents the degree of influence of each candidate keyword on the current behavioral disturbance. The keyword influence score vector is read and a dynamic truncation threshold filtering strategy is applied. Candidate keywords with scores higher than the threshold are encapsulated into a set to obtain the final set of keywords to be updated, which is used to guide the fine-tuning of local model parameters.

8. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 7, characterized in that, The step of calculating the behavior pattern disturbance event trigger signal based on the single-window behavior activity entropy value and the personalized entropy threshold includes: Obtain the single-window behavior activity entropy value of the current window and the corresponding personalized entropy threshold, and use numerical comparison logic to perform size judgment processing on the two to obtain the behavior pattern disturbance event trigger signal that represents whether the behavior pattern has undergone drastic evolution.

9. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 7, characterized in that, The process of calculating and processing the candidate keyword adjacency matrix based on the behavioral pattern perturbation event trigger signal includes: The system receives the behavior pattern disturbance event trigger signal as an activation instruction, performs semantic coupling degree calculation based on the statistical results of high-frequency new words and long-tail words within the sliding time window, and obtains a candidate keyword adjacency matrix representing strong correlation between words.

10. The potential keyword mining and prediction method based on time-series search behavior analysis according to claim 7, characterized in that, The step of obtaining a set of behavior-coupled paths by path matching based on the adjacency matrix of candidate keywords includes: The system obtains recent click-through path sequence data of users, and uses a graph traversal method to perform cross-class jump path matching processing on the click-through path sequence and the adjacency matrix of the candidate keywords to extract a set of behavioral coupling paths that reflect the instantaneous shift trajectory of user interests.

Images